I build models, agents, and the software around them.
I'm Yash, a final-year AI/ML student at BIT Mesra. I work on physiological sensing, LLMs, and agents, alongside the APIs, interfaces, and infrastructure that turn experiments into working applications.
Résumé · LinkedIn · Email · LeetCode
01 · PRISM
Physics-aware model selection for contactless heart-rate estimation.
A team research project at BIT Mesra. A 10-dimensional optical and motion representation feeds a meta-regressor that predicts specialist rPPG errors and selects the lowest-error model. Kalman-stabilized face regions, a FastAPI video pipeline, remote inference, and a React analysis interface connect the research to an application.
Source & architecture · Reported experiments · Web demo
Tool-using agents with policy enforced at the write boundary.
A LangGraph agent handles order questions and refunds; application code rechecks identity, ownership, and refund policy before committing a decision. PostgreSQL row locks and a unique approval index guard against duplicate refunds. Redis coordination, streamed audit events, model fallback, and Opik traces support the deployed application.
The committed Gemini baseline covers 82 cases, with all four deterministic metrics at 100%; this describes the recorded evaluation.
Try the app · Architecture · Evaluation record
Long contracts, limited GPU memory, and an inspectable model lifecycle.
A contract-analysis pipeline adapts Hermes-4-14B to classify 17 ContractNLI hypotheses and return supporting clauses. The training recipe combines QLoRA, 4-bit NF4, BF16, gradient checkpointing, and PyTorch DDP. MLRun artifacts, versioned adapters, and SageMaker/vLLM serving code extend the work beyond training. Historical evaluation and serving measurements are documented with their scope and limitations.
Training & evaluation · Lifecycle & implementation status
04 · TaskForge CloudOps
An independently built application, from interface to infrastructure.
I built the task-management application and its engineering implementation end to end: Next.js, a NestJS API, PostgreSQL, and a dedicated BullMQ worker. Docker Compose, Terraform-provisioned AWS infrastructure, and GitHub Actions deliver digest-pinned releases. Background reports expose real queue state; verification covers worker recovery, persistence, and deployment rollback in CI.
Application & architecture · Background jobs · Delivery pipeline
05 · AstraDrive
A perception-to-control loop in autonomous-driving simulation.
A simulation pipeline connects BeamNG.tech camera and radar inputs to hybrid CV/UFLDv2 lane perception, YOLO detection, and PID steering. It includes cruise control and a separate radar/TTC braking mode. The active loop runs perception locally; HTTP aggregation and Docker services are experimental alternatives. Recorded demos show individual experiments in this research prototype.
Architecture & scope · Demos & project site
- Models & experiments — Python, PyTorch, OpenCV, Transformers, PEFT/QLoRA.
- Applications & data — TypeScript, React/Next.js, FastAPI, NestJS, PostgreSQL, Redis.
- Delivery & operations — Docker, GitHub Actions, AWS, Terraform, Google Cloud Run.
Interested in AI/ML engineering, agent reliability, applied research, and backend systems. If you're working on a problem in these areas, let's talk.


