Software engineer who happens to speak fluent machine learning.
I build software end to end. Some of it thinks — fine-tuned models, retrieval, agents that stay on a leash. All of it ships — React and FastAPI up top, Docker and EC2 underneath.
LoRA fine-tuning, retrieval, guardrails, output validation. I'd rather have a model that stays inside its lane than one that demos well.
React and TypeScript on the front, Node or FastAPI behind it, Mongo and Redis where they earn their keep. Interfaces people actually use.
Docker, Nginx, EC2, GitHub Actions. I deploy what I build and I keep it alive afterwards.
A small language model that knows one subject properly. LLaMA and Qwen, fine-tuned with LoRA on a curriculum dataset built through retrieval - so a college can run its own tutor on cheap hardware and no student data leaves the building.
The chapter's official platform, and the one I get paged about. Registrations, content management, and the operations behind every technical event we run. I wrote it, containerised it, put it on EC2 behind Nginx, and I still maintain it.
Paste a news article, get it back split into claims with a credibility score on each. Agents go and find evidence on the open web and argue it out. React for the verdicts, MongoDB caching so the second run is fast.
Ask the Bhagavad Gita a question and get an answer that can show its working. Verses are embedded with Sentence Transformers into a FAISS index; the closest ones come back first and Llama 3 answers from those, not from memory. Public, and it stays up.
Sign the alphabet at a webcam and read it back as text, live. YOLO finds the letter, MediaPipe tracks the hand, and a feature and inference layer I wrote on top ties the two together. 0.99+ precision, 0.90 mAP, and quick enough to keep up with a real hand.
Remote (Hong Kong)
Took apart the best agentic harnesses out there - MCP tooling, context management, orchestration - and wrote up what the team should copy and what it shouldn't.
Hybrid
Built an agentic pipeline that writes database algorithms without ever seeing the raw data, and worked on what a RAG system should do when its sources disagree.
Remote
Rebuilt the company site in React with a reusable component set and AVIF assets. Noticeably faster on every device.
Vice President of UPES-CSA, a 50+ member technical chapter — after two years running its engineering team. I've shipped software with people, not just alone.
PyTorch
Transformers
LoRA / PEFT
RAG
FAISS
OpenCV
Python
TypeScript
JavaScript
Java
React
Next.js
Node / Express
FastAPI
Tailwind
SQL
Bash
MongoDB
MySQL
Redis
Vector DBs
Docker
AWS EC2
Nginx
GitHub Actions
LinuxThat's the job description, not the insult.
hover a spine ↑ — the three that actually changed how I work.
Straight answers. I'd rather show you the repo than a slide about it.
Calm on deploy day - the checklist exists so nobody has to be brave.
Habits over heroics. Small, consistent shipping - yes, the book worked.
Pick one. The other three come along anyway.