I write code. I make it work. I make it stay working.

$ whoami
> software engineer & applied ai.
> currently: shipping small models, building production systems.
> stack: python, fastapi, pytorch, next.js, postgres.
> focus: llm fine-tuning (lora/dpo), rag, mlops.

find me: github · huggingface · linkedin · email

4-stage fine-tuning pipeline (SFT → Reward-SFT → DPO → LoRA) on Qwen2.5-1.5B-Instruct. VRAM from 24GB to 8GB via LoRA, gradient checkpointing, bfloat16. Exported to GGUF Q8_0 (1.53GB) for CPU-only inference via llama.cpp. Deployed as FastAPI server with SSE streaming. 510+ downloads on HuggingFace. [PyTorch] [PEFT] [TRL] [llama.cpp] [FastAPI] [Docker]

Full-stack AI hiring platform (student, recruiter, admin) as a 3-service monorepo. Groq-powered resume parsing cuts manual shortlisting by 60% across 200+ resumes. JWT + Supabase RLS auth, GitHub Actions CI with RBAC smoke tests. Deployed on Vercel + Render with zero downtime. Won 2nd prize at Innovision (out of 30+ teams) and offered pre-seed funding. [Next.js 15] [Express] [FastAPI] [Supabase] [Docker]

Document Q&A with per-user FAISS index isolation and Celery for background chunking. Sub-2s retrieval latency on 100k+ token corpora. JWT refresh-token rotation with Redis-backed revocation. [LangChain] [FAISS] [OpenAI] [Celery] [Redis]

RandomForest with SMOTE oversampling and 25 engineered features. ROC-AUC 0.94. Redis prediction cache, MLflow experiment tracking, drift detection, and weekly retrain logic. Full stack via docker-compose. [Scikit-Learn] [MLflow] [PostgreSQL] [Redis] [Docker]

FastAPI-based resume screener using Groq (llama-3.1-8b-instant). JWT auth, per-user screening history, admin role, statistics dashboard. SQLite + SQLAlchemy. [FastAPI] [Groq] [SQLAlchemy] [JWT]

Next.js 15 + React 19 web app for text + speech translation between languages. Uses free endpoints (MyMemory, LibreTranslate, Lingva) with fallback, voice input/output, dark mode, and translation history in localStorage. [Next.js] [React] [TypeScript] [Tailwind]