AI/ML engineer · LLM fine-tuning · multi-agent systems · production backends
Final-year B.E. AI & Data Science student at SRCOE, Pune. I fine-tune small language models, build agents that run on modest hardware, and ship FastAPI backends that hold up outside a notebook.
📍 Pune, India · 🔎 Open to AI/ML and backend roles (internship or full-time), Pune or remote
| 🧠 Model development | LoRA fine-tuning of small LLMs, training-data curation, quantization (GGUF) and local deployment with Ollama |
| 🤖 Agents | Multi-agent orchestration with LangGraph, tool-calling, RAG memory, and agents that run offline on small local models |
| ⚙️ Backends | FastAPI, PostgreSQL, Redis, Celery, JWT/RBAC, Docker, deployed services |
| 🖥️ Frontends | React 18, TypeScript, Next.js, Tailwind, Three.js |
I trained two LoRA models myself using Unsloth, and exported them for local inference.
Health, skincare and nutrition assistant · Qwen3.5-0.8B
- Goal: a very small model that stays useful at low memory, for skincare, nutrition and general wellness questions (weight loss, weight gain).
- Training data: JSONL generated and curated on purpose, skipping examples the base model already handles well, so training time goes into what the model is actually weak at.
- Output: merged 16-bit weights plus a
Q4_K_MGGUF that runs through Ollama on consumer hardware.
Constitution of India Q&A · Qwen2.5-1.7B
- LoRA fine-tune on Constitution question-answer data, trained with Unsloth.
Along the way: I followed a RAG → LoRA → RAFT learning path, built a local RAG pipeline (Ollama, FAISS, EmbeddingGemma), and run my models locally through Ollama.
local-code-agent · private repo An offline coding agent built for a small local model (Qwen 0.8B Q4 on a 2 GB GPU, standard-library Python only). The harness does the planning and the model only handles tiny, well-defined jobs: write the tests, write the code, fix the failing output. Every step runs tests first, retries within a limit, rolls back on failure, and commits on success. It has built working apps step by step, including an HRMS and a task manager.
MAP, Multi-Agent Platform
Planner → Executor → Analyzer → Memory pipeline on LangGraph, with FAISS RAG memory, a fallback engine with circuit breakers, RS256 JWT with RBAC, and Celery workers. I led a team of four through the build using a strict branch → PR → review workflow.
LangGraph FastAPI Celery PostgreSQL Redis FAISS React 18 TypeScript
OmClaw Local-first autonomous coding agent on Ollama with tool-calling and Docker-sandboxed execution.
Skill Scout · private repo, in progress An AI-assisted job hunter. It discovers jobs from public sources (Greenhouse, Lever, company careers pages), scores them with explainable matching rules, keeps a manual-apply queue, and tracks applications so nothing is submitted twice. FastAPI, SQLite, Ollama, with a guided web UI and CI.
SRS, Automated Customer Support Resolution · frontend Ticket classification, a confidence-based decision engine that auto-resolves or escalates to a human, an OpenAI fallback chain, and an admin dashboard. Deployed on Render and Vercel.
Deepfake Detection API EfficientNet-B4 classifier served through FastAPI with GradCAM explanations for each prediction, plus a React frontend.
College Event System Clubs, events, RSVP and waitlists, QR attendance, PDF certificates, and real-time notifications.
Reimbursement System Multi-tenant expense platform with OCR receipt parsing and approval workflows. Built in 8 hours as an Odoo × VIT Pune Hackathon 2026 finalist.
DAB AI · private repo, in progress A voice-agent builder platform on FastAPI, PostgreSQL and Next.js.
More on my portfolio (React, Three.js, GSAP).
AI/ML Intern · Kemuri Technology · Dec 2025 – Jun 2026 Worked on COE AI, a production image-classification service for marine garbage: GPT-4o mini vision, OCR and weight estimation, integrated with a live Kuroco CMS. Tasks came through tickets and went through senior code review before merge.
Want to work together or talk models and agents? Reach me on LinkedIn.