I’m Biprayan — an AI/ML engineer focused on multi-agent architectures, local-first LLM deployments, and production-grade RAG pipelines. I design systems where autonomous agents reason, verify, and act — with a bias for things that actually run.
I’m a B.S. student of Artificial Intelligence & Data Science at IIT Guwahati, expected to graduate in 2028. I work on the parts of AI that are the hardest to demo and the most rewarding to ship: multi-agent orchestration, safety critics that catch hallucinations before they reach a user, and retrieval pipelines that hold up on messy, real-world data.
My work leans local-first — I’d rather run a small model on the user’s machine than phone home to a paid API. That constraint is the reason I got into agentic systems in the first place, and it shapes every architecture decision I make.
I’m looking for internships and full-time roles where I can keep doing exactly that — design, build, and ship AI systems that are reliable enough to put in front of people who depend on them.
Each one started as a real problem — a hallucinated clinical answer, a video transcript no one could query, a blog draft that took an hour. The list below is the work I’m most proud of because each one shipped.
A short list of the things I’ve shipped more than once and can be trusted to run with. If your team needs any of this, I’m a fast ramp.
State-graph agents that fan out, verify, and re-evaluate. LangGraph architectures with safety-critic nodes, Pydantic-typed contracts between agents, and structured handoffs. I’ve shipped agents that pass evaluation harnesses, not just demos.
Air-gapped, zero-per-query-cost deployments on CPU and consumer GPUs. Quantization, Ollama, faster-whisper, GGUF, model selection. I treat cloud APIs as a fallback, not the default.
Vector search that actually finds the right passage, with citations a reviewer can verify. FAISS, sentence-transformers, Cohere, hybrid lexical+dense retrieval, evaluation against ground-truth chunks. I write the pipeline, the eval, and the guardrails.
From notebook to something a non-engineer can use. FastAPI services, Streamlit apps, Pydantic schemas, structured logging, and reproducible experiments. I default to writing the boring infra first so the interesting parts stay reliable. LangGraph architectures with safety-critic nodes, Pydantic-typed contracts between agents, and structured handoffs. I’ve shipped agents that pass evaluation harnesses, not just demos.
A living list. I learn new stacks quickly when a project demands it; these are the ones I have real depth in.
The academic and external context for the work above.
Indian Institute of Technology Guwahati
Coursework spans machine learning, deep learning, statistical modelling, optimization, and systems. Outside class, I lead independent builds in agentic AI and contribute to open-source tooling around LangGraph.
Self-directed · client & research work
Designing and shipping production AI systems — clinical decision support, autonomous content pipelines, RAG over video corpora, local-first companion agents. Each one starts from a real problem and ships with evaluation.
I read every message myself. The fastest way in is email; LinkedIn or GitHub also work.
I’m currently taking on internships and full-time roles starting 2026, plus select research collaborations. If you’re working on agentic systems, applied LLMs, or local-first deployment — I’d like to hear about it.
Or use the form below — it opens your email client with the message pre-filled. The form is wired to biprayanc@gmail.com.