Carrinson

Projects

AI - Customer Support Assistant

• Built a Telegram bot that answers customer questions from a 50-page PDF knowledge base, returning a grounded answer in under 3 seconds.
• Moved LLM calls off the request path with Celery workers over Redis, so the bot acknowledges each message in milliseconds and serves concurrent conversations without blocking.
• Implemented the RAG pipeline end-to-end: PDF chunking, embedding, Chroma DB vector retrieval, and prompt assembly for a Groq-hosted model.
• Built a JavaFX admin dashboard backed by SQLite for managing knowledge-base context and reviewing bot responses.

Github Link Blog Link

AI Healthcare Assistant & Disease Prediction Platform

• Designed and built the full backend: FastAPI service on PostgreSQL with Alembic migrations and JWT authentication (register / login / me) working end-to-end against a self-managed Ubuntu VPS.
• Shipped three ML-backed prediction endpoints (symptom triage, diabetes risk, heart risk) fronted by a red-flag safety layer that escalates emergency presentations before any model output is returned.
• Served self-hosted Hugging Face models on my own infrastructure rather than a hosted AI API, and provisioned the box myself — Docker, ufw, fail2ban, key-only SSH.
• Set up GitHub Actions CI to run the pytest suite on every push, covering the auth flow and the safety layer.
• Built the two clients against the same API: a React Native (Expo Router) patient app and a React + Vite clinician dashboard.

Github Link

AI Request Validator API

• Schema-first validation service for LLM API requests: Pydantic v2 enums, field validators, and typed response models reject malformed payloads before they reach the model, preventing whole classes of downstream failure.
• Separated service and router layers behind typed contracts, so validation rules change without touching transport code.

Github Link