# psr.ai

> Prabhjot Singh Rai builds production AI systems for domains where a wrong output is expensive.

- Site: https://www.psr.ai
- Email: hi@psr.ai
- X: https://x.com/psr_ai

Judgment over generation. Models are cheap. Taste, evals, and architecture are not.

## Path

Process research in Grenoble, France. Research at the University of Minnesota, Minneapolis. Then production software and models.

Now: Platform Technical Lead, AI/ML at ISHI Health.

Education: Stanford University (AI graduate certificate), University of Minnesota (M.S. Computer Science), IIT Roorkee (B.Tech. Engineering).

## How AI is used

AI is used to cover more of the operation: search, prototypes, implementations, tests. The standard stays with the engineer. Unreviewed model output is not shipped into high-consequence paths. Fluency is not treated as evidence.

## Selected systems

- Custom LLM through targeted fine-tuning at Nophin (YC W22): 100% accuracy on the defined evaluation set, cloud-agnostic inference with Ray Serve and Modal.
- Real-time industrial control at Phaidra: 5,000 points/second, ~30 GB/day, 20% higher operating efficiency at 99.9% uptime.
- Emergency-response driving (ADAS-E), AWS-funded: 90% better emergency-response performance in CARLA. https://www.psr.ai/projects/adas-e

## Project notes

- [RL Fine-Tuning of Language Models with GenRM-CoT](https://www.psr.ai/projects/genrm-cot)
- [PocketSheet: memory-augmented reasoning](https://www.psr.ai/projects/pocketsheet)
- [Brain Age Prediction Using Deep CNNs](https://www.psr.ai/projects/brain-age)
- [ADAS for Emergency Services](https://www.psr.ai/projects/adas-e)
- [Sleep Detection](https://www.psr.ai/projects/sleep-detection)
- [OpenAI Gym Lunar Lander](https://www.psr.ai/projects/lunar-lander)

## Optional

- Projects index: https://www.psr.ai/projects
- RSS: https://www.psr.ai/feed.xml
- Contact: https://www.psr.ai/contact
- Meet: https://www.psr.ai/meet