# agentsite-static-version=eb04c841-c4a3-4e69-888c-73241139c95e # perit.ai > AI-optimized mirror of perit.ai containing 190 pages totalling 190,998 words of clean markdown content, structured data, and semantic HTML. Original source: https://perit.ai. Last updated: 2026-11-16T12:48:08.350Z. Each page is available as HTML (with JSON-LD structured data) and Markdown (text-only, ideal for LLMs and RAG). ## Homepage - [Perit — frontier training data, RL environments and benchmarks, graded by people who do the work](/content/site-root.html): Perit collects and annotates frontier training data — speech, video, image and text — runs speech-to-speech RL environments and benchmarks, and captures embodied data for robotics. 4,000+ hours delivered by 650 transcribers and aligners across 9 locales. Operated by Zyno AI Inc, backed by Y Combinator. (1,076 words) ## Articles & Blog Posts - [blogs/banking-agent-procedure-final-state-check/index.html](/content/blogs/banking-agent-procedure-final-state-check/index.html) (3,091 words) - [blogs/hinglish-code-switching-speech-recognition/index.html](/content/blogs/hinglish-code-switching-speech-recognition/index.html) (4,101 words) - [research/benchmark/speech-to-speech/index.html](/content/research/benchmark/speech-to-speech/index.html) (538 words) - [Conversation Samples — Perit](/content/conversations/index.html): The dataset primarily consists of natural, spontaneous, and unscripted conversations involving 2 to 5 participants. Each speaker is recorded on a separate, isolated audio channel, ensuring complete voice separation without mixing speakers into a single waveform. Conversations maintain clear turn-taking patterns, with overlapping speech accounting for less than 1% of the recordings. (487 words) - [Data annotation — Perit](/content/services/data-annotation/index.html): The Foundry bench: 650 transcribers and aligners, a written guide, a calibration test nobody skips, and QA sampling scored against WER. Audio delivered; image, video and text on the same bench. (1,215 words) - [Perit Blog - Frontier training data, collected and graded by people who do the work.](/content/blogs/index.html): Frontier training data, collected and graded by people who do the work. (638 words) - [Marketplace · Perit Expert Tasks](/content/expert-tasks/marketplace/index.html): Containerised agentic tasks with a measured pass rate each, and a per-model breakdown of what every one of them discriminated against. (1,115 words) - [Egocentric · Physical Intelligence — Perit](/content/physical-intelligence/egocentric/index.html): Monocular first-person video from a chest or head mount, recorded by people doing everyday tasks in the rooms where those tasks actually happen — for imitation learning, hand–object understanding and embodied pretraining. (985 words) - [Speech-to-text leaderboard · PERIT-STT — Perit](/content/research/benchmark/speech-to-text/leaderboard/index.html): 22 speech-to-text models run on the same real-world audio and scored word by word against transcripts written and reviewed by people. (371 words) - [Data collection — Perit](/content/services/data-collection/index.html): Speech collected at scale — 4,000+ hours, nine locales, consent on every file — and video, image and text collected to spec on the same recruiting and QA machinery. (831 words) - [Find the money in a payout run that reported clean · Perit Expert Tasks](/content/expert-tasks/marketplace/payout-dispatch/index.html): Nine merchants' payouts left their available balances on Sunday night and nothing arrived. The run reported 240 payouts and no errors, the morning reconciliation came back clean, and the bank's statement says it sent everything it was given, which is true. Work out what happened from the incident folder. (331 words) - [Report every file a reference library holds · Perit Expert Tasks](/content/expert-tasks/marketplace/bib-library-guard/index.html): Two bibliography reports went out wrong and both went out silently: a file a reference manager had exported in the wrong encoding was dropped, and nothing in the report said a record was missing. Report what the library holds, including the files that cannot be read. (329 words) - [Benchmarks · Perit Expert Tasks](/content/expert-tasks/benchmarks/index.html): Seventeen capabilities measured in containers, each with its own artifact contract, base image and resource ceiling. (727 words) - [Close a replay gap in a TOTP verify path · Perit Expert Tasks](/content/expert-tasks/marketplace/pyotp-replay-guard-skew-window/index.html): A login flow verifies TOTP codes statelessly, so the same six digits submitted twice inside one step both pass. Make a code work once across every account, while the store of accepted codes stays bounded on a box that authenticates far more accounts than it ever sees twice. (335 words) - [Take a leaver out of the team tables · Perit Expert Tasks](/content/expert-tasks/marketplace/identity-not-position/index.html): An account leaving the company has to come out of the team tables. The rules were written over four years, they are dated, and the handbook says which of its own sections carry force and which are one person's notes. (331 words) - [Compose SSH command batches from a maintenance plan · Perit Expert Tasks](/content/expert-tasks/marketplace/remote-batch-compose/index.html): A fleet tool turns a maintenance plan into the command batches it sends to one server over SSH. Compose them against a written spec, from a plan and a runbook that are supplied separately, so that sending a batch twice is not the same as doing the work twice. (340 words) - [Plan which MCP tools a turn can offer · Perit Expert Tasks](/content/expert-tasks/marketplace/mcp-tool-budget/index.html): A turn can hand a model only so many tools. Built-in servers take part of a fixed budget and the configured servers share what is left. Decide which servers fit, given that a server which does not fit should not cost the turn the ones that would have. (329 words) - [Scoring · Perit Expert Tasks](/content/expert-tasks/scoring/index.html): How one container run becomes a number: pass@k, reward written by a verifier, and the six gates a task passes before a score against it counts. (538 words) - [Withhold tax against a schedule only a service holds · Perit Expert Tasks](/content/expert-tasks/marketplace/tds-rate-service/index.html): Compute the tax to be deducted at source on a company's vendor payments. The rate schedule is not in the repository at all: an MCP server is the only authority for rates, thresholds and the amendments that moved them, so every figure has to be fetched rather than assumed. (319 words) - [Decide what a command string would actually run · Perit Expert Tasks](/content/expert-tasks/marketplace/bash-exec-resolve/index.html): A command-approval hook sits in front of a shell and has to answer, before anything runs, which external programs a command string would start and whether any of them is destructive. The hook it replaces reads the string as text and keeps being got around by wrappers put in front of the real command. (326 words) - [Data marketplace — Perit](/content/marketplace/index.html): 29 datasets Perit can license, across 10 categories: speech in 122 languages, egocentric and general video, production codebases, agentic RL environments, field documents and clinical cohorts. Every listing carries its volume, formats, licence status and availability. (1,152 words) - [Decide which test results are evidence about production · Perit Expert Tasks](/content/expert-tasks/marketplace/backend-disagreement/index.html): A nightly suite runs against a test backend that is not the production one, and the two disagree in ways somebody has documented. Decide which of the suite's results say something about production and which say something only about the test backend. (333 words) - [Recover a session's context figure from its transcript · Perit Expert Tasks](/content/expert-tasks/marketplace/transcript-context-probe/index.html): A compaction reminder is silent in long sessions and premature in short ones, and both faults are in the probe that reads a session's context size. The transcript is appended to while the session runs, it is read from the end rather than whole, and the usage block it carries reports the same cache tokens in two shapes. (339 words) - [Reconcile a month's input tax credit · Perit Expert Tasks](/content/expert-tasks/marketplace/gst-itc-reconcile/index.html): A finance team closes a credit run by comparing what suppliers filed against what the company itself recorded, giving each invoice a status and an amount. The rules are the company's own and they are written down across several files, not all of which are current. (338 words) - [Report who holds a resource's data after a chain of hand-offs · Perit Expert Tasks](/content/expert-tasks/marketplace/normalise-then-collide/index.html): A fleet of resources is installed, handed between teams and sometimes reinstalled, and each holds data one team is custodian of at a time. What a hand-off does and what a reinstall does to that custody were written down at different times. (333 words) - [Hand Pose · Physical Intelligence — Perit](/content/physical-intelligence/hand-pose/index.html): Egocentric capture paired with wrist cameras and marker gloves, so the delivered hand pose is measured rather than estimated — including the keypoints the head camera loses behind an object or the other hand. (993 words) - [Merge a coding agent's queued messages into one turn · Perit Expert Tasks](/content/expert-tasks/marketplace/queue-turn-start/index.html): A desktop app runs coding agents behind a per-agent message queue. Messages that waited together are still read apart, so one thought costs four turns and pays for the whole session four times. Merge what waited together, in the order it was written, without the queue losing its liveness when a turn does not start. (336 words) - [Evidence · Perit Expert Tasks](/content/expert-tasks/evidence/index.html): Every trial is kept, not summarised: the bundle, the image digests, the trajectories and the verifier's own output, beside the number. (332 words) - [Decide which imported fields were altered on the way in · Perit Expert Tasks](/content/expert-tasks/marketplace/import-exactness/index.html): An export of stored records was imported back into the system and the import path changed some of them in transit. Work out which fields were altered and which were always like that, against a handbook written over several years and dated accordingly. (334 words) - [Depth samples · Physical Intelligence — Perit](/content/samples/depth/index.html): 3 sample depth episodes to watch and download — RGB with an aligned per-frame depth map from stereo headsets, LiDAR phones and head-mounted depth cameras. (362 words) - [Rebuild a settlement ledger against withholding mints · Perit Expert Tasks](/content/expert-tasks/marketplace/curve-quote-settle/index.html): A venue settles trades against quote pools, and the ledger has to be rebuilt from state so it can be reconciled against the chain. Some quote assets are tokens whose mints withhold part of every transfer and do not publish their terms, so the terms have to be inferred from the transfers that were made and what was kept out of them. (348 words) - [Sample datasets · Physical Intelligence — Perit](/content/samples/index.html): Watch and download sample episodes of egocentric, depth, gripper and teleoperation capture, see what every download contains, and the hardware each line is recorded with. Hand-pose, tactile, BYOR and annotation samples on request. (625 words) - [Reconcile an inbox against responses that arrive out of order · Perit Expert Tasks](/content/expert-tasks/marketplace/inbox-revision-fence/index.html): A desktop inbox shows a change the moment somebody makes it, while the request is still in flight, and responses come back in whatever order the network gives them. The list has twice jumped backwards: an older response restored a finished row, and a change that failed stayed on screen as though it had worked. (332 words) - [UMI Gripper samples · Physical Intelligence — Perit](/content/samples/umi-gripper/index.html): 2 sample umi gripper episodes to watch and download — Handheld wrist-camera gripper demonstrations with a 6-DoF trajectory and jaw width per frame. (305 words) - [Video Annotation · Physical Intelligence — Perit](/content/physical-intelligence/video-annotation/index.html): Egocentric, teleoperation, robot and policy-rollout video annotated on Foundry: sub-task boundaries, natural-language captions, hand and object masks, success and failure labels — to a written guide, behind a calibration test, with agreement reported per batch. (984 words) - [Open datasets — Perit](/content/datasets/index.html): 5.1 hours of full-duplex, one-speaker-per-channel English support conversations, free on Hugging Face under CC-BY-4.0. The remaining 15.5 hours on request. (425 words) - [Trust & security · Perit Expert Tasks](/content/expert-tasks/intake/index.html): An uploaded bundle is a stranger's code. The first five intake stages read it, measure it and refuse it without executing a single byte, and that is a property we test. (355 words) - [Hospital EMR + Radiology Cohort (India) — Perit](/content/marketplace/indian-hospital-emr-radiology/index.html): A de-identified hospital cohort: clinical notes, radiology studies and their reports. 15.6M patients; 78.7M notes; 13.4M radiology studies. Rights pending (de-identification unverified), lead (metadata only). (252 words) - [AI Data Annotation (FTC) — Perit](/content/careers/ai-data-annotation-ftc-india/index.html): Full-time · India (Remote) · $260–365 / month. The same transcription work on a full-time footing, for people who want the hours rather than the flexibility. (332 words) - [Organisational Supply: 14 Company Codebases + Ops Data — Perit](/content/marketplace/organisational-supply-company-codebases/index.html): Codebases and operational data from 14 companies across conversational AI, health, gaming, ed-tech, travel and HR-tech. ~580 repos, ~16.7M LOC; ops data est. up to hundreds of TB (email, Slack, Jira, DBs, call recordings, video). Rights pending, lead (metadata only). (336 words) - [Teleoperation · Physical Intelligence — Perit](/content/physical-intelligence/teleoperation/index.html): Trained operators drive leader–follower arms or VR-controlled robots through repeatable tasks. Every episode logs synchronised camera streams with joint states and actions — the data a policy is actually trained on. (946 words) - [Freelance Audio Conversation Rater — Perit](/content/careers/freelance-audio-conversation-rater/index.html): Freelance · Bengaluru, Karnataka, India (Remote) · $210–315 / month. The same rating work, freelance — 30 to 50 hours a week, on your own schedule. (366 words) - [PERIT-STT — Perit](/content/research/benchmark/speech-to-text/index.html): 22 speech models ranked on word error rate (wer), with every supporting metric and latency. PERIT-STT scores models on real support recordings — mobile lines, crosstalk, accents and the domain vocabulary that decides whether a transcript is usable downstream. (755 words) - [For authors · Perit Expert Tasks](/content/expert-tasks/for-authors/index.html): A task only earns its place if models fail it honestly. The pipeline measures that for you and says exactly why a submission did not qualify, with the file and the line. (129 words) - [Privacy — Perit](/content/privacy/index.html): What Perit collects from contributors, annotators, applicants and customers across speech, video, image and text; how recordings and footage are consented to on the file; where they live; and how to withdraw. (1,312 words) - [Spanish Freelance AI Data Annotator (Transcription) — Perit](/content/careers/spanish-freelance-transcription/index.html): Freelance · France (Remote) · Rate on application. Transcribe and check Spanish audio against the guideline. Work any hours, from anywhere. (265 words) - [For labs · Perit Expert Tasks](/content/expert-tasks/for-labs/index.html): You get the score and everything behind it: the bundle, the image digests, every trial that ran, and the verdict at each gate. Not a CSV of prompts and a number you have to take on trust. (133 words) - [FAQ — Perit](/content/faq/index.html): How fast Perit starts, what a quote contains, which modalities, locales and hardware we collect with, where robotics teams get real-world training data, how consent and security work, and how to join the bench. (1,051 words) - [Enterprise Workflow Data (Slack / Email / Tickets) — Perit](/content/marketplace/enterprise-workflow-data/index.html): Scrubbed workplace operations data: Slack threads, email and tickets, delivered as JSONL. To spec. Rights pending, collecting. (250 words) - [UMI Gripper · Physical Intelligence — Perit](/content/physical-intelligence/umi-gripper/index.html): Handheld UMI-style grippers with a wrist-mounted fisheye camera let people demonstrate grasps anywhere — kitchens, shops, workshops — without a robot in the room. Manipulation data with the diversity a lab cannot stage. (941 words) - [Bring Your Own Rig · Physical Intelligence — Perit](/content/physical-intelligence/bring-your-own-rig/index.html): Robotics teams with a proprietary capture system, a custom helmet, a new gripper or a robot of their own get the whole bench — recruiting, environments, trained operators, SOPs and QC — without handing over the stack. (1,023 words) - [Agentic RL Tasks (Verified Environments) — Perit](/content/marketplace/verified-agentic-rl-tasks/index.html): Verified agentic tasks that each ship with the environment, a reference solution, hidden graders and frontier-model calibration. 7 tasks; 7 archives, ~44.6 MB total. Rights pending (licence not yet recorded), lead (metadata only). (276 words) - [Multilingual Social-Media App Codebase (Full Org) — Perit](/content/marketplace/multilingual-social-app-codebase/index.html): A full consumer social-media org: Android, iOS, PWA, backend, data engineering and ML, with 14 years of history. 388 repos; 4.2M programming LOC / 54M tokens; 285,870 commits (2012-2026). Rights pending, lead (metadata only). (339 words) - [Insurance Underwriting Agent (Inland Marine) RL Tasks — Perit](/content/marketplace/inland-marine-underwriting-rl-tasks/index.html): Scrubbed production episodes of a seven-step, tool-using underwriting agent, packaged as RL tasks with ground truth and a programmatic grader. 12 episodes; 12 tasks with ground truth; 25 preference pairs; 84 step labels. Perit-built (licence approval confirmed), sample available. (380 words) - [Word alignment leaderboard · PERIT-ALIGN — Perit](/content/research/benchmark/word-alignment/leaderboard/index.html): 14 word-alignment models run on the same real-world audio and scored word by word against timestamps marked and reviewed by people. (790 words) - [Logistics & E-commerce Production Code Excerpts — Perit](/content/marketplace/logistics-ecommerce-code-excerpts/index.html): Production code excerpts from logistics and e-commerce systems: billing and GST, COD, TAT and scheduling. 8 excerpts / 1,443 LOC sample; up to 50 repos claimed. Rights pending, lead (metadata only). (361 words) - [Real-World Field Images (Documents, Parcels, Proof-of-Delivery) — Perit](/content/marketplace/field-image-documents-deliveries/index.html): Field photos of documents, parcels and proof-of-delivery from 10+ countries, 2020 to 2025. ~2.85M images (2020-2025, 10+ countries). Rights review, sample on request. (247 words) - [Sign up · Perit Expert Tasks](/content/expert-tasks/signup/index.html): Create an author account on Perit. (9 words) - [Startup Pitch Decks & Investor Memoranda — Perit](/content/marketplace/startup-pitch-decks/index.html): Consumer and D2C startup pitch decks and investor memoranda, 18 to 50 slides each. 24 PDFs sample (18-50 slides); 10,000+ claimed. Rights pending, lead (metadata only). (275 words) - [US-English Customer Support Calls (Channel-Separated) — Perit](/content/marketplace/us-english-support-calls-stereo/index.html): Inbound US support calls with the agent on the left channel and the customer on the right. ~19 hrs. Rights review, sample on request. (297 words) - [Agentic & Long-Horizon Benchmark Tasks — Perit](/content/marketplace/agentic-long-horizon-benchmarks/index.html): Long-horizon agent tasks with environments, programmatic graders and reference runs, built by Perit. 11 task families. Perit-built, sample available. (323 words) - [Physical Intelligence — Perit](/content/physical-intelligence/index.html): Real-world training data for robotics, humanoids and VLA models: egocentric video, depth, teleoperation, gripper demonstrations, hand-pose ground truth, tactile streams, capture on your own rig and video annotation — recorded in real homes, shops, workshops and worksites across nine locales, quoted in 48 hours. (739 words) - [Online-Gaming Platform Production Backends — Perit](/content/marketplace/online-gaming-platform-backends/index.html): Five production backends from an online-gaming platform: payments, contests and game services. 5 repos, ~203K LOC, ~19.5K commits. Rights pending, lead (metadata only). (254 words) - [Licensed English Podcast Catalog (Multitrack) — Perit](/content/marketplace/licensed-english-podcast-multitrack/index.html): Long-form English podcast conversations with a separate WAV track per speaker and a mixed M4A. 1,000 hr catalog; 3-min preview. Licence in negotiation, on request. (282 words) - [Request data — Perit](/content/request-data/index.html): Ask for the data your model is missing — speech, video, image or text. Speech delivered at scale; delivery spec and licence terms published. (1,192 words) - [Multilingual Speech Library (122 Languages) — Perit](/content/marketplace/multilingual-speech-122-languages/index.html): A licensed speech library in 122 languages, about 13.8K hours of it transcribed. ~2.89M hrs. Licensed catalog (sublicense), on request. (299 words) - [Trust & Security — Perit](/content/trust/index.html): How Perit handles customer data: encryption, role-scoped access, graders who stream rather than copy, consent on every file, region pinning and deletion. SOC 2, ISO 27001, ISO 27701 and TPN Gold; GDPR, CCPA and HIPAA compliant. (625 words) - [RL Environments — Perit](/content/rl-environments/index.html): Speech-to-speech RL environments for frontier labs and voice-agent teams: live calls with trained speakers who interrupt, change their mind and hand over the phone, scored on whether the call ended well. Live in ten sectors. (763 words) - [Education · PERIT-STT — Perit](/content/research/benchmark/speech-to-text/education/index.html): Admissions, fees and course guidance, often with a parent and a student on the same line. (222 words) - [Tactile samples · Physical Intelligence — Perit](/content/samples/tactile/index.html): Tactile samples on request — Video with a synchronised per-taxel pressure stream and contact events per finger or jaw. (238 words) - [Research — Perit](/content/research/index.html): Perit Research: two speech benchmarks — transcription and word alignment — with every model run on the same audio under the same scoring, plus the method and field notes behind them. (561 words) - [Retail & e-commerce · PERIT-STT — Perit](/content/research/benchmark/speech-to-text/retail/index.html): Order status, returns and refunds, with catalogue names that no general model has seen. (226 words) - [Exam-Prep Lecture Video (Civil Services & Engineering Entrance) — Perit](/content/marketplace/exam-prep-lecture-video/index.html): Full-length exam-prep lectures for civil services and engineering entrance, recorded at 1080p. 2 full lectures (2.4 h) sample. Teacher permission (verbal); written pending, sample on request. (177 words) - [Video Annotation samples · Physical Intelligence — Perit](/content/samples/video-annotation/index.html): Video Annotation samples on request — Segments, captions, masks and success labels per episode, in JSON or your schema. (180 words) - [Terms of service — Perit](/content/terms/index.html): The agreement covering the Perit site, Foundry and the contributor workspace, and everything we deliver: data collection and annotation across speech, video, image and text, licensed datasets, speech benchmarks, RL environments and embodied-data pilots. (1,206 words) - [Perit Expert Tasks: a task is not data until it has been graded](/content/expert-tasks/index.html): Perit Expert Tasks builds and grades containerised tasks for frontier labs. Six gates, then a person who has read the transcript. (697 words) - [Depth · Physical Intelligence — Perit](/content/physical-intelligence/depth/index.html): Egocentric capture with a depth channel — stereo headsets, LiDAR-equipped phones or custom stereo rigs — so a model can learn where things are, not only what they look like. (1,098 words) - [Careers — Perit](/content/careers/index.html): Open roles on Foundry, the Perit annotation bench — transcription and word alignment today, with image, video and text work on the same bench — fully remote, pay published on every listing. (640 words) - [Method — Perit](/content/research/method/index.html): How a Perit benchmark run is executed: held-out audio, two-pass references, adjudication, bootstrapped intervals and a working human control. (577 words) - [Teleoperation samples · Physical Intelligence — Perit](/content/samples/teleoperation/index.html): 2 sample teleoperation episodes to watch and download — Leader–follower and VR teleoperation episodes with synchronised cameras and per-step state and action logs. (319 words) - [Tactile · Physical Intelligence — Perit](/content/physical-intelligence/tactile/index.html): Tactile gloves and sensorised grippers log contact and pressure per fingertip alongside video, so a model can learn the difference between holding an egg and holding a mug from the signal, not the pixels. (936 words) - [Logistics & delivery · PERIT-STT — Perit](/content/research/benchmark/speech-to-text/logistics/index.html): Driver and dispatch calls from vehicles, warehouses and doorsteps. Half of it is hands-free. (234 words) - [Healthcare · PERIT-STT — Perit](/content/research/benchmark/speech-to-text/healthcare/index.html): Appointment intake, triage lines and medication questions, dense with drug names and dosages. (239 words) - [Travel & hospitality · PERIT-STT — Perit](/content/research/benchmark/speech-to-text/travel/index.html): Rebookings during disruption — airport noise, urgency, and three changes in one turn. (241 words) - [Banking · PERIT-STT — Perit](/content/research/benchmark/speech-to-text/banking/index.html): Card disputes, KYC checks and payment failures, where a misheard digit is the whole error. (224 words) - [PERIT-ALIGN — Perit](/content/research/benchmark/word-alignment/index.html): 14 speech models ranked on timing accuracy @100 ms, with every supporting metric and latency. PERIT-ALIGN scores the word-level timestamps a speech model returns — where each word starts and where it ends — against reference timings, so captions, redaction and audio search land on the right moment. (337 words) - [Record & earn — Perit](/content/record/index.html): Get paid to record your voice. Twenty-minute sessions on your phone, $20 per approved audio hour plus locale bonuses, paid within seven days. (553 words) - [Telecom · PERIT-STT — Perit](/content/research/benchmark/speech-to-text/telecom/index.html): Plan changes, SIM swaps and outage triage, mostly over mobile lines that drop mid-sentence. (235 words) - [Hand Pose samples · Physical Intelligence — Perit](/content/samples/hand-pose/index.html): Hand Pose samples on request — Head and wrist video with 21 measured keypoints per hand per frame, occlusions included. (225 words) - [Perit Duplex Support Conversations (English) — Perit](/content/marketplace/perit-duplex-support-conversations/index.html): Open full-duplex English support role-play: healthcare scheduling, billing, travel, insurance and IT. Open 5.1 h / 45 convos; full 15.5 h / 180 convos. Open (CC-BY-4.0), ready now. (306 words) - [General Video Library (75 Content Categories) — Perit](/content/marketplace/general-video-library/index.html): A broad licensed video library across 75 content categories, from sport and transport to food and professions. ~4.49M hrs; 136.8M videos. Licensed catalog (sublicense), on request. (268 words) - [Field support & utilities · PERIT-STT — Perit](/content/research/benchmark/speech-to-text/field-support/index.html): Technicians on rooftops and in basements, talking through a fix with a machine running. (227 words) - [AI Data Annotation (Fixed Term) — Perit](/content/careers/ai-data-annotation-fixed-term/index.html): Contract · India (Remote) · $260–365 / month. Listen to audio, check the transcript against what was actually said, and apply the guideline consistently. (375 words) - [Multilingual Two-Speaker Conversations — Perit](/content/marketplace/multilingual-two-speaker-conversations/index.html): Two-speaker natural conversations recorded in 13 languages, about an hour of sample audio per language. ~1 hr per language across 13 languages (sample set). Rights review, sample on request. (312 words) - [Scanned Personal & Financial Documents (OCR) — Perit](/content/marketplace/scanned-personal-documents/index.html): Scanned personal and financial documents for OCR and document understanding, in English and Indian languages. ~450 TB claimed. Rights pending (consent unverified), lead (metadata only). (274 words) - [Licensed Egocentric Activity Video — Perit](/content/marketplace/licensed-egocentric-activity-video/index.html): Licensed first-person video of 29 skilled activities, including clothing, art, furniture, cleaning and weaving. 55,306 hrs; 1.1M files. Licensed catalog, on request. (299 words) - [Pricing — Perit](/content/pricing/index.html): Programs priced on scope — locale, annotation depth, turnaround, exclusivity and rig — with a written quote in 48 hours and contributor pay as its own line. Contact for pricing. (546 words) - [AI Data Annotation (Fixed Term Contract) — Perit](/content/careers/ai-data-annotation-ftc-bengaluru/index.html): Contract · Bengaluru, Karnataka, India (Remote) · $260–365 / month. Fixed-term transcription work on the Bengaluru posting, fully remote. (380 words) - [Studio Voice-Actor Conversations (Mono) — Perit](/content/marketplace/voice-actor-simulated-conversations/index.html): Simulated conversations performed by consented studio voice actors, recorded mono. Not stated. Consented performers (own recording), on request. (257 words) - [Bring Your Own Rig samples · Physical Intelligence — Perit](/content/samples/bring-your-own-rig/index.html): Bring Your Own Rig samples on request — Your hardware and format, our crews, rooms and QC — under NDA. (246 words) - [Household Egocentric Video + 23k Task Taxonomy — Perit](/content/marketplace/household-egocentric-video/index.html): First-person household task video, labelled against a 23,000-task taxonomy. Faces blurred. Up to 70k hrs; sample 16 clips (~9.9 GB); 23k-task list. Rights review, sample on request. (331 words) - [Indic Multilingual Speech (Redacted) — Perit](/content/marketplace/indic-multilingual-speech-redacted/index.html): Conversational speech across seven Indic languages and English, with personal details redacted. 38 MP3 sample clips. Rights review, sample on request. (258 words) - [Audio Conversation Rater (Fixed Contract) — Perit](/content/careers/audio-conversation-rater-fixed-contract/index.html): Contract · Bengaluru, Karnataka, India (Remote) · $260 / month. Rate what an AI voice agent said and how it sounded, and defend every score in your own words. (320 words) - [Chinese Freelance AI Data Annotator (Transcription) — Perit](/content/careers/chinese-freelance-transcription/index.html): Freelance · Singapore · Malaysia (Remote) · $45 / audio hour. Transcribe and check Chinese audio against the guideline, paid per hour of audio you deliver. (252 words) - [Freelance AI Data Annotation (Alignment) — Perit](/content/careers/freelance-ai-data-alignment/index.html): Freelance · Bengaluru, Karnataka, India (Remote) · $17–31 / audio hour. Put every spoken word where it belongs in the timeline. No fixed schedule, no weekly minimum. (288 words) - [Data residency — Perit](/content/data-residency/index.html): Where Perit stores and processes audio, video, images and labels — Frankfurt, Northern Virginia or Mumbai — how a collection is pinned to one region, and how transfers and deletion are handled. (479 words) - [AI Data Annotation (Remote) — Perit](/content/careers/ai-data-annotation-remote-bengaluru/index.html): Contract · Bengaluru, Karnataka, India (Remote) · $260–365 / month. Transcription review, fully remote, open to candidates based around Bengaluru. (338 words) - [B2B Tech Sales Cold Calls (Indian English) — Perit](/content/marketplace/b2b-sales-cold-calls/index.html): Unscripted outbound sales calls in Indian English, recorded over telephony, in three length buckets. 90 calls; ~4.6 hrs unique, in three length buckets. Rights pending (no consent or licence proof), lead (metadata only). (285 words) - [German Freelance AI Data Annotator (Transcription) — Perit](/content/careers/german-freelance-transcription/index.html): Freelance · Switzerland · Austria (Remote) · Rate on application. Transcribe and check German audio against the guideline. Work any hours, from anywhere. (268 words) - [Image Annotation & Verification (Custom) — Perit](/content/marketplace/image-annotation-verification/index.html): Image annotation and verification collected and labelled by Perit to your spec. To spec. Perit-collected, collecting. (313 words) - [Public services · PERIT-STT — Perit](/content/research/benchmark/speech-to-text/public-services/index.html): Benefits, permits and civic helplines, where callers read reference numbers aloud, badly. (225 words) - [Human-Skills Graded Scenario Q&A — Perit](/content/marketplace/human-skills-graded-qa/index.html): Scenario questions on judgement, EQ, communication and leadership, with graded answers and mentor interviews. 246K-400K+ tagged answers; 182 mentor voice interviews. Rights pending (eval only today), lead (metadata only). (249 words) - [Hindi Freelance AI Data Annotator (Transcription) — Perit](/content/careers/hindi-freelance-transcription/index.html): Freelance · India (Remote) · Rate on application. Transcribe and check Hindi audio against the guideline. Work any hours, from anywhere. (284 words) - [Mechanical & Sheet-Metal CAD Projects — Perit](/content/marketplace/mechanical-cad-sheet-metal/index.html): SolidWorks and STEP projects for industrial machinery, HVAC cabinets and frames, with drawings and BOMs. 164 files across sample batches; ~300 TB claimed. Rights pending (client release needed), lead (metadata only). (283 words) - [Insurance · PERIT-STT — Perit](/content/research/benchmark/speech-to-text/insurance/index.html): First-notice-of-loss calls taken from roadsides and hospital corridors, rarely from a quiet room. (224 words) - [Indie Mobile Game Catalogue (Unity, C#) — Perit](/content/marketplace/unity-mobile-game-catalogue/index.html): A defunct studio's full catalogue of Android games in Unity: driving, racing, shooters, hunting, runners and board games, with source, scenes, prefabs and art. 38 project archives, ~61.3 GB compressed. Rights pending (developer-owned, unverified in writing), lead (metadata only). (266 words) - [Tech-Bootcamp Lecture & Mock-Interview Video — Perit](/content/marketplace/edtech-bootcamp-video/index.html): Recorded coding classes, lessons and mock interviews from a tech bootcamp, in English and Hinglish. 3-file sample 3.37 h; ~123K hrs claimed. Rights pending, lead (metadata only). (162 words) - [Field notes — Perit](/content/research/notes/index.html): Notes on how we build and score speech benchmarks: rubric design, held-out audio, human controls, and why task success beats latency. (120 words) - [WorldDraft: The Umbrella Opened in the Wrong Colour - Perit Blog](/content/blogs/worlddraft-video-models-4d-world-source-leakage/index.html): WorldDraft asks whether video models can add realism without rewriting an executable 4D world. The umbrella opens correctly—and keeps the wrong colours. (2,818 words, 2026-11-15) - [Gemini Robotics 2.5 Hit 96%. What Was the Denominator? - Perit Blog](/content/blogs/gemini-robotics-2-5-mercedes-96-percent-success/index.html): DeepMind reports 96% success for Apollo 2 on Mercedes-Benz lines. Here is what the denominator, resets and recovery reveal about the result. (2,903 words, 2026-11-15) - [One Mug, Three Teachers: Egocentric Video vs Teleoperation vs UMI - Perit Blog](/content/blogs/egocentric-video-teleoperation-umi-robot-training-data.html): Egocentric video, teleoperation and UMI-style grippers can record the same task but preserve different evidence. Here is how robotics teams should choose—and combine—the three forms of robot training data. (2,358 words, 2026-11-15) - [AlphaProof Code: The AI Wrote 6.5 Million Lines. Humans Had to Trust 26,000. - Perit Blog](/content/blogs/alpha-proof-code-signal-formal-verification/index.html): AlphaProof Code generated 6.5 million lines of machine-checked proof for Signal. The deeper story is why specification—not code—may become the scarce enterprise artifact. (2,561 words, 2026-11-13) - [MKA: The Agent Was Given a Photograph. It Turned It Into ASCII. - Perit Blog](/content/blogs/multimodal-knowledge-agent-mka-visual-reasoning/index.html): An AI agent receives a road photograph, converts it into ASCII and misses the pothole. What MKA reveals about visual knowledge work, compute and expert evaluation. (2,562 words, 2026-11-13) - [The AI Read Every Update. It Still Used the Wrong State. - Perit Blog](/content/blogs/ai-agent-state-tracking-statebench/index.html): An AI reads every warehouse update, cites a real one, and still sends the driver to the wrong bay. StateBench shows why retrieval is not the same as knowing what is true now. (3,034 words, 2026-11-10) - [Weight-Space Programming: The Model Knew the Answer. It Couldn't Stop. - Perit Blog](/content/blogs/weight-space-programming-llm-edit-neural-network-weights.html): Weight-space programming lets an LLM edit a neural network's weights without training. What GPT-6 Astra's 0.016% Qwen edit fixed, and what it did not. (1,755 words, 2026-11-05) - [Can AI Close the Loop Between Discovery and Experiment? - Perit Blog](/content/blogs/can-ai-close-the-loop-between-discovery-and-experiment.html): Scientific discovery has always involved a difficult exchange between what researchers think might happen and what actually happens in the laboratory. An idea may look convincing in a mathematical model,… (2,390 words, 2026-11-05) - [The Rise of Materials Designed by Machines - Perit Blog](/content/blogs/the-rise-of-materials-designed-by-machines/index.html): What if scientists could describe the properties they need and let artificial intelligence help design the material itself? For decades, materials discovery has relied on experimentation, computational simulations, and repeated… (2,420 words, 2026-11-04) - [When Proteins Become a Search Space for AI - Perit Blog](/content/blogs/when-proteins-become-a-search-space-for-ai/index.html): For billions of years, nature has been building proteins. Some carry oxygen through our blood, others help digest food, recognize invading pathogens or transmit signals between cells. Almost every process… (4,262 words, 2026-11-04) - [AI Is Beginning to Discover What Science Hasn't Found Yet - Perit Blog](/content/blogs/ai-is-beginning-to-discover-what-science-hasnt-found-yet.html): Artificial intelligence is moving beyond predicting outcomes and analyzing existing research. It is beginning to help scientists formulate new hypotheses, connect overlooked findings and explore questions that were once difficult… (3,213 words, 2026-11-03) - [Small Enough to Run Locally, Smart Enough to Go Multimodal - Perit Blog](/content/blogs/small-enough-to-run-locally-smart-enough-to-go-multimodal.html): Artificial intelligence has spent years moving in one direction. Models have become larger, computing requirements have increased, and infrastructure has grown more expensive. Every new generation has reinforced the idea… (3,016 words, 2026-10-29) - [Gemini 4 Carbon Is Built to Break the Cybersecurity Bottleneck - Perit Blog](/content/blogs/gemini-4-carbon-is-built-to-break-the-cybersecurity-bottleneck.html): Finding a security vulnerability is one thing. Understanding whether someone can exploit it, figuring out why it exists, and fixing it without breaking something else is a much harder job.… (4,057 words, 2026-10-29) - [The Rise of Persistent AI Agents - Perit Blog](/content/blogs/the-rise-of-persistent-ai-agents/index.html): Imagine you have a report to finish before tomorrow morning. You need to research several websites, compare the findings, update a spreadsheet, and turn everything into a polished document. You… (1,660 words, 2026-10-26) - [Block Dessa: Speech-Only AI and What Conversation Teaches It](/content/blogs/block-dessa-speech-only-ai-conversation-model/index.html): Block says Dessa learned conversation from speech alone. Explore its origins, release status, and what voice AI must prove beyond natural delivery. (3,239 words, 2026-10-24) - [The Missing Audit Trail of Artificial Intelligence - Perit Blog](/content/blogs/the-missing-audit-trail-of-artificial-intelligence/index.html): Imagine a researcher asking an AI assistant a straightforward question about a scientific discovery. The model responds confidently, names a researcher, describes an experiment, and explains the result. The answer… (3,384 words, 2026-10-22) - [The Agent Remembered the Wrong Lesson - Perit Blog](/content/blogs/ai-agent-memory-wrong-lessons-error-reproduction/index.html): An AI agent can turn one mistaken interpretation into reusable guidance. New research shows how memory reproduces errors, and what enterprise systems should preserve instead. (2,945 words, 2026-10-21) - [The Security Boundary AI Agents Keep Crossing - Perit Blog](/content/blogs/the-security-boundary-ai-agents-keep-crossing/index.html):  Why sandboxing, network isolation, and credential boundaries matter when AI agents execute real-world actions. The Moment an Agent Crosses the Line An AI agent becomes fundamentally different from a conventional… (2,741 words, 2026-10-18) - [Rethinking AI Inference with Maia 300 - Perit Blog](/content/blogs/rethinking-ai-inference-with-maia-300/index.html): Microsoft Maia 300 reflects a growing shift in AI infrastructure. Companies are no longer focused only on building more powerful AI models. They are also trying to make those models… (2,288 words, 2026-10-16) - [Why Is Apple Teaching an AI to Read Pictures of Text? - Perit Blog](/content/blogs/apple-lensvlm-ai-read-pictures-of-text/index.html): Apple’s LensVLM turns long documents into miniature page images, then expands only what matters. What 4.3× compression really buys and costs. (3,111 words, 2026-10-14) - [An Hour of Audio in 13 Seconds? What Local Voice AI Still Has to Learn - Perit Blog](/content/blogs/embedding-gemma-2-local-voice-ai-transcriber/index.html): EmbeddingGemma becomes a local transcriber. What its 13-second demo reveals about word fidelity, small specialists, and enterprise voice AI. (3,083 words, 2026-10-13) - [The New AI Workspace - Perit Blog](/content/blogs/the-new-ai-workspace/index.html): A model can write a function, explain a research paper, or summarize a spreadsheet in seconds. But the work rarely ends there. The code still needs to run, the findings… (3,287 words, 2026-10-10) - [Voice Deepfake Detection Fails on Phone Calls. Here's Why - Perit Blog](/content/blogs/voice-deepfake-detection-phone-audio/index.html): A cloned voice helped move €95m out of an Italian bank. Why voice deepfake detection fails on real phone calls, and the data a detector actually needs. (3,620 words, 2026-10-07) - [Six AI Labs Shipped Decision Models in 17 Days - Perit Blog](/content/blogs/decision-models-launches-2026/index.html): Six AI labs shipped decision models in 17 days. How Jev and Cloudflare Clef return probabilities, not text, what they cost, and where they fail. (3,182 words, 2026-10-07) - [Synthetic Data Can Scale AI. But Can It Teach Reality? - Perit Blog](/content/blogs/synthetic-data-can-scale-ai-but-can-it-teach-reality.html): The Synthetic Data Scaling Problem Synthetic data offers something that is difficult to achieve with real-world data: scale. Instead of waiting for enough real examples to appear, teams can generate… (3,228 words, 2026-10-07) - [The Robot Didn’t Fail. The Dataset Did. - Perit Blog](/content/blogs/the-robot-didnt-fail-the-dataset-did/index.html): When More Data Still Isn’t Enough A warehouse robot can sort the same type of package thousands of times and still struggle when the shelf is rearranged, the lighting changes,… (2,803 words, 2026-10-06) - [pass^k: Measuring AI Agent Reliability Across Repeated Calls - Perit Blog](/content/blogs/pass-k-ai-agent-reliability/index.html): pass^k measures whether an AI agent succeeds on every repeated call: how to estimate it without bias, grade on backend state, and budget human calls. (1,838 words, 2026-10-05) - [When AI Gets It Right Most of the Time: The Problem With Rare Cases - Perit Blog](/content/blogs/when-ai-gets-it-right-most-of-the-time-the-problem-with-rare-cases.html): The Data Distribution Is Not the Real World A dataset does not become representative just because it is large. A model can be trained on millions of examples and still… (1,573 words, 2026-10-03) - [Engineering Reliable AI Agents for Production Systems - Perit Blog](/content/blogs/engineering-reliable-ai-agents-for-production-systems.html): Why AI Agents Become Unreliable at Scale An AI agent can complete a task correctly in a test environment. It finds the right information, calls the appropriate tool, and returns… (3,369 words, 2026-10-02) - [AI Is Running Out of Data: What Happens When the Internet Stops Scaling? - Perit Blog](/content/blogs/ai-is-running-out-of-data-what-happens-when-the-internet-stops-scaling.html): The next AI bottleneck may not be compute For the last few years, the recipe for building better AI has looked fairly straightforward: make the models bigger, give them more… (2,739 words, 2026-10-01) - [The Benchmark Nobody Sees - Perit Blog](/content/blogs/the-benchmark-nobody-sees/index.html): A model goes from 94% to 97%. How test leakage can inflate that gain, how contamination happens by accident, and how to keep a benchmark held out. (1,349 words, 2026-09-30) - [Word Timestamps And Forced Alignment for Voice Agents - Perit Blog](/content/blogs/word-timestamps-forced-alignment-for-voice-agents/index.html): How transcription choices change word timestamps, where forced alignment fails on conversation, and what timing benchmarks reveal about voice agents. (3,454 words, 2026-09-30) - [Why More Call Recordings Won't Fix Your Voice Agent - Perit Blog](/content/blogs/rl-environments-voice-agents/index.html): A recorded call shows one outcome. An RL environment lets a voice agent try other decisions, with real evidence to check whether practice transfers. (4,775 words, 2026-09-30) - [From Raw Data to AI-Ready Datasets: The Complete Data Annotation and Collection Pipeline - Perit Blog](/content/blogs/from-raw-data-to-ai-ready-datasets-the-complete-data-annotation-and-collection-pipeline.html): Why AI-Ready Data Requires More Than Annotation Building an AI model does not start with choosing a model architecture. It starts with getting the right data. Images, videos, audio recordings,… (3,530 words, 2026-09-30) - [Teleoperation Explained - Perit Blog](/content/blogs/teleoperation-explained/index.html): What a teleoperation episode records, how its streams stay synchronised, and why calibration and session metadata matter for robot learning. (1,332 words, 2026-09-30) - [Order 5% First: How a Data Collection Pilot Works - Perit Blog](/content/blogs/pilot-economics/index.html): A 5% pilot tests whether a data brief works before the full order runs. The cost arithmetic, what to inspect, and why QA continues after the pilot. (1,550 words, 2026-09-29) - [Engineering Quality in AI Data: Combining Human Expertise with Automated QA - Perit Blog](/content/blogs/engineering-quality-in-ai-data-combining-human-expertise-with-automated-qa.html): Where Does Quality Break Down in AI Data? AI models depend heavily on the quality of the data used to train them. But as datasets grow to thousands or millions… (1,681 words, 2026-09-26) - [What "95% Accuracy" Means in Data Annotation QA - Perit Blog](/content/blogs/what-95-accuracy-means-in-data-annotation-qa/index.html): What QA-sample accuracy, word error rate, agreement and rework rate each tell you about an annotation batch. (1,555 words, 2026-09-24) - [From Human Demonstrations to Self-Improving Robots: The Data Behind Physical AI - Perit Blog](/content/blogs/from-human-demonstrations-to-self-improving-robots-the-data-behind-physical-ai.html): How Robots Learn from Data Many robot skills start as a set of recordings, and one of the most direct ways to learn from them is by demonstration. A person… (2,432 words, 2026-09-24) - [Training Data for Humanoid Robots: What Physical AI Needs to Learn From the Real World - Perit Blog](/content/blogs/training-data-for-humanoid-robots-what-physical-ai-needs-to-learn-from-the-real-world.html): Why Humanoid Robots Need a Different Kind of Training Data Humanoid robots are learning to work in environments designed for people, but teaching them to do so requires more than… (1,872 words, 2026-09-22) - [Collection, Annotation or Both? - Perit Blog](/content/blogs/collection-annotation-or-both/index.html): Does a model failure need new inputs, better labels or a clearer spec? A guide to choosing data collection, annotation or both. (2,751 words, 2026-09-20) - [Turning Real Conversations Into AI Training Data - Perit Blog](/content/blogs/turning-real-conversations-into-ai-training-data/index.html): How structured human conversation gets collected, transcribed and graded into a dataset a model can actually learn from. (253 words, 2026-09-16) - [Turning Real Medical Conversations into AI Training Data - Perit Blog](/content/blogs/turning-real-medical-conversations-into-ai-training-data.html): The Data Hiding in Every Doctor’s Visit There’s a version of every doctor’s visit that never makes it into the medical record. The way a patient struggles to describe a… (2,136 words, 2026-09-15) - [The Model Got 8/10. For What? - Perit Blog](/content/blogs/the-model-got-8-10-for-what/index.html): A refund answer earns eight out of ten, but no refund was submitted. How RL rubrics define success, combine criteria and shape what an agent learns. (5,745 words, 2026-09-13) - [The AI Lab, the Vortex, and the 10,000-Agent Weekend - Perit Blog](/content/blogs/the-ai-lab-the-vortex-and-the-10000-agent-weekend/index.html): OpenAI’s proposed forced-fluid proof, the search involving 10,000 agents, and the questions formal verification leaves about scientific credit. (2,623 words, 2026-09-11) - [How Fine-Tuned Open Models Are Quietly Beating the Generalists at Their Own Game - Perit Blog](/content/blogs/how-fine-tuned-open-models-are-quietly-beating-the-generalists-at-their-own-game.html): There is an old saying that “jack of all trades, master of none, yet oftentimes better than a master of one”. In artificial intelligence, this saying is starting to break.… (966 words, 2026-09-07) - [Meta's New AI Doesn't Just Transcribe:It Actually Listens - Perit Blog](/content/blogs/metas-new-ai-doesnt-just-transcribeit-actually-listens.html): There’s a moment in almost every recorded conversation where the transcript falls apart,someone talks over someone else, a technical term gets butchered, or the software just can’t keepup with how… (952 words, 2026-09-06) - [The Hardest Word in Voice AI Might Be “Yeah” - Perit Blog](/content/blogs/the-hardest-word-in-voice-ai-might-be-yeah/index.html): “Yeah” can acknowledge a speaker or interrupt them. What TurnBench measures, where voice agents struggle, and how anticipation changes response timing. (2,053 words, 2026-09-06) - [Why the score is task success, not latency — Perit](/content/research/notes/task-success-not-latency/index.html): Latency is the easiest speech number to publish and the easiest to game. It belongs in the report, not at the top of it. (116 words, 2026-08-14) - [A rubric a stranger can apply — Perit](/content/research/notes/rubric-a-stranger-can-apply/index.html): If two reviewers grading the same call disagree, the criterion is broken — not the reviewer. (112 words, 2026-07-30) - [What held-out means when the data is audio — Perit](/content/research/notes/held-out-audio/index.html): Text benchmarks leak by being quoted. Audio leaks by being reused — the same speaker, the same room, the same script. (105 words, 2026-07-08) - [The human control is not a ceiling — Perit](/content/research/notes/human-control-is-not-a-ceiling/index.html): Publishing a human line at 100 makes the benchmark useless the moment a model gets close. (119 words, 2026-06-19) ## Listings & Categories - [AI Engineering Archives - Perit Blog](/content/blogs/tag/ai-engineering/index.html) (202 words) - [LLM Archives - Perit Blog](/content/blogs/tag/llm/index.html) (166 words) - [AI Agents Archives - Perit Blog](/content/blogs/tag/ai-agents/index.html) (167 words) - [Research Archives - Page 2 of 3 - Perit Blog](/content/blogs/category/research/page/2/index.html): Benchmarks, methodology write-ups and findings from Perit AI’s research team. (360 words) - [Artificial Intelligence Archives - Page 3 of 4 - Perit Blog](/content/blogs/category/artificial-intelligence/page/3/index.html) (274 words) - [Research Archives - Page 3 of 3 - Perit Blog](/content/blogs/category/research/page/3/index.html): Benchmarks, methodology write-ups and findings from Perit AI’s research team. (146 words) - [Tech Update Archives - Page 2 of 2 - Perit Blog](/content/blogs/category/tech-update/page/2/index.html) (39 words) - [Artificial Intelligence Archives - Page 2 of 4 - Perit Blog](/content/blogs/category/artificial-intelligence/page/2/index.html) (268 words) - [Artificial Intelligence Archives - Page 4 of 4 - Perit Blog](/content/blogs/category/artificial-intelligence/page/4/index.html) (168 words) - [Research Archives - Perit Blog](/content/blogs/category/research/index.html): Benchmarks, methodology write-ups and findings from Perit AI’s research team. (349 words) - [Tech Update Archives - Perit Blog](/content/blogs/category/tech-update/index.html) (288 words) - [Robotics Archives - Perit Blog](/content/blogs/category/robotics/index.html) (161 words) - [Featured Archives - Perit Blog](/content/blogs/category/featured/index.html) (90 words) - [Guides Archives - Perit Blog](/content/blogs/category/guides/index.html): Practical how-tos for teams sourcing and annotating frontier training data. (73 words) - [Physical Intelligence Archives - Perit Blog](/content/blogs/category/physical-intelligence/index.html): Notes from the field on egocentric video, teleoperation and robot training data. (207 words) - [Artificial Intelligence Archives - Perit Blog](/content/blogs/category/artificial-intelligence/index.html) (329 words) ## About Pages - [Contact · Perit Expert Tasks](/content/expert-tasks/contact/index.html): Tell us what your model gets wrong. Benchmarking, dataset access and custom evaluation, answered by the engineer who owns the area. (119 words) - [About — Perit](/content/about/index.html): Perit is a training-data company built on one idea: the person who does the job is the right person to record it, transcribe it and grade it. Operated by Zyno AI Inc, backed by Y Combinator. (664 words) - [Contact — Perit](/content/contact/index.html): One form for a collection brief, a failing clip, an annotation sample, a physical-AI pilot, a benchmark entry, a security questionnaire or a role — read by a person, answered within 48 hours. (107 words) ## Resources - [Full Page Index](/index.html): Browse all cached pages with rich metadata - [About This Cache](/about.html): Methodology, technical details, and usage guidelines - [XML Sitemap](/sitemap.xml): Machine-readable sitemap for crawler discovery - [Robots.txt](/robots.txt): Crawler directives - [Agentsite Network](https://agentsite.network/network.html): Public index of Agentsites and their machine-readable resources