
Perit helps AI labs get the data and training environments they need to build models that work in the real world.
We collect, annotate, and deliver high-quality datasets across audio, video, image, and code. We also build RL environments and benchmarks for models that need to interact with real people and real-world conditions.
We have a network of 16,000+ trained experts across different sectors and 12+ locales.
Within 4 months of launch, we have delivered 6,000+ hours of speech data across 9 locales, 10k+ hours of egocentric data, and 100s of RL tasks for major AI labs, already running multi-million-dollar contracts (targeting $10M by Demo Day🤞).
Need data your model is failing on? We collect it.
Speech, egocentric video, hand pose, depth, and other multimodal datasets can be collected from real people using our own devices, with consent on every file.
Models need to be evaluated in situations that are messy and unpredictable, not just on static test sets. We currently have live speech to speech RL environments & benchmarks across 10 sectors, where trained operators interact with models like real customers. They interrupt, change their requests, provide incomplete information, have noisy connections, and behave like people actually do.
The environments measure task success, latency, interruption recovery, repair of missing or incorrect information, instruction adherence, and tone. Failed interactions come back with the audio and annotations, not just a score.
We also build custom RL environments for new use cases.
We build expert-authored coding datasets and agentic RL environments across software engineering, systems, security, data, and scientific computing. Models work inside sandboxes on terminal, repository, web, and UI tasks, then receive deterministic rewards from hidden tests, reference outputs, and strict execution contracts.
Each environment is calibrated against frontier models and includes adversarial cases that expose plausible but incorrect solutions. Labs receive the failed runs and traces needed to improve performance on messy, long-horizon engineering work, not just a score.
https://www.youtube.com/watch?v=wQwCInBdrCI
Our superpower is speed and accuracy: quickly spinning up workflows and mobilizing experts.
Perit was founded by @Kshitij Dixit and @Poorva Dixit, both IIT Kharagpur alumni and second-time YC founders.
I previously scaled @Zeo Route Planner to millions of courier drivers across 100+ countries, generating millions of dollars in revenue in just a few years.
Poorva is the cofounder and CTO. She has previously trained models at Google, built scalable systems at Atlassian, and built map-quality systems at Uber. Both of them reached the same conclusion: models still can't perform on many real-world tasks, and with the right specialized datasets their performance can be quickly improved.
If you are an AI lab or company looking for specialized training data, annotation, or RL environments, we would love to talk.
Email: kshitij@perit.ai
Website: https://perit.ai