SAMEERPANDEY

Running MediaPipe's landmark detection directly inside FastAPI's async event loop caused CPU starvation. Every frame stalled concurrent network connections, causing socket timeouts and jitter.

Offloaded model inference to dedicated worker processes. Redis Pub/Sub serves as the high-throughput message bus: inference workers publish landmark coordinates while FastAPI WebSocket subscribers broadcast frames to connected clients.

Separated CPU-bound compute from I/O-bound networking. Concurrent WebSocket clients receive frame coordinates with non-blocking event loops, with session persistence in PostgreSQL and Nginx handling reverse proxying.

Manufacturing deviation capture in pharma QMS relied on slow, manual form entry with zero support for multi-format document uploads.

Built a structured processing pipeline with intent routing, LLM extraction via Groq, Pydantic validation, and automatic SQLite/PostgreSQL fallback.

Structured deviation records from text, chat, or uploaded documents (PDF, DOCX, XLSX, TXT, scanned images) with merge/diff logic.

Commercial nutrition planners assume Western food datasets and high budgets (₹800+/day). Standard LLM generation hallucinated impossible caloric density and ignored local price realities for Indian students.

Created an optimization pipeline where dual LLMs propose Indian ingredient combinations under hard daily spending caps (e.g. ₹80–₹120/day). Every macro proposal is strictly parsed by Pydantic and verified against authentic food composition APIs before presentation.

Eliminated macro hallucinations and delivered actionable, budget-constrained diet regimes. Automatic fallback between primary and secondary LLM APIs ensures request completion without client-side retries.

Standard RL reward functions optimizing solely for calories triggered reward hacking: consuming pure oil and sugar at every step.

Redesigned the reward function with penalties for allergen usage, budget violations, and dietary variety floors across 50 Indian ingredients.

Agent converged on balanced, realistic meal sequences with three difficulty tiers. Deployed to Hugging Face with a live Gradio UI so users and evaluators can interact with the environment.