Building AI Systems
Bangalore, India
I am looking to join a strong team as a software engineer. My preference is a fast moving, small team, somewhere between 1 and 100 people, building something people actually want. I can join immediately. You can reach me at nikm5502@gmail.com.
I love tinkering and keep building things on the side, sometimes just to learn, sometimes to see how the technology is moving.
In my previous roles I worked across the full stack, including production deployments on AWS and building pipelines. I also spent time on diffusion and fine tuning, including writing architectures in PyTorch. Since then I have focused on building AI systems.
Right now I am building Mini-pi, a coding agent from scratch, to strip away the abstractions these frameworks hide. Turns out simple patterns win.
| Languages | Python, TypeScript, C/C++ |
|---|---|
| AI | Agents (LangChain, Phoenix), FAISS, PyTorch, RL (OpenEnv, GRPO), VAEs, Latent DiT |
| Backend | Django, FastAPI, PostgreSQL, MongoDB, Redis, Celery |
| DevOps | AWS (EC2, ECS, S3, Lambda, Load Balancer), Terraform, Docker, Linux |
Code artifacts from multiple projects
A coding agent built from scratch, with sandboxed tools, session persistence and branching, context compaction, hooks and extensions.
It grew out of my tinkering with harnesses on SWE-bench Verified, with OpenTelemetry tracing and LLM as a judge trajectory evals.
A key value store written in C. Concurrent TCP server, RESP parsing, key expiry, and master replica synchronization.
Trained a VAE from scratch to generate and interpolate hand drawn sketches, then wrote up how the training and inference actually work.
Built the dataset by scraping sites, running OCR on scanned books and transcribing YouTube audio, generated question answer pairs with the Gemini API, then fine tuned Gemma-3-1B-IT on rented GPUs.
A LangGraph agent that asks clarifying questions first, then scrapes Amazon, Flipkart and Google Search to compare products and return the top three matches.
The official platform for campus elections at IIT Roorkee, where 10K+ students voted inside a two hour window. Tamper resistant voting, and the AWS infrastructure that held through the peak.
Built for the Indian Knowledge Systems division at the Ministry of Education. Upload a certificate template, point it at a spreadsheet, and it generates personalised certificates and emails them out, with SQS and Lambda doing the work in the background.
An RL environment where an LLM agent coordinates hospital supply chains across wards and enterprise systems, scored by a deterministic reward formula with no LLM judge. Trained Qwen3.5-2B against the live simulation with GRPO.
Paste a YouTube link, get a finished vertical reaction reel. An agent reads the transcript, picks the sharpest moment, writes the commentary and calls the Runway APIs to generate every asset.
A GenAI app that segments tumors from MRI scans and writes diagnostic reports with treatment plans.
Led technical projects end to end, and ran workshops, lectures and mentoring for 100+ students across the club.
How I built minipi, a coding agent heavily inspired by pi. The blog is language agnostic and covers the agent loop, talking to the model, tools and the sandbox, session persistence and branching, compaction, steering, slash commands, hooks, extensions and the TUI.
A runnable script walked through line by line to show what distributed training actually does: synchronization, seeding, and how the GPUs talk to each other.
A step by step breakdown of the inference pipeline, from the T5 encoder through the latent DiT to the VAE decoder.
V8, libuv and the event loop, and why the event driven architecture is built the way it is.
Every file is the same binary digits. What decides whether they are an image, audio or text is only how you agree to read them.