Skip to content
Roshan Chhetri

Roshan Chhetri · AI Engineer

I build production LLM systems, not demos.

Most of my work is agentic and retrieval-based. The part I care about is the gap between a demo that works once and a system that holds up with real users on it.

Open to relocation, requires visa sponsorshipImmediate joiner, open to relocate within India· Sikkim, India
Roshan Chhetri, AI Engineer

Selected work

All projects

An agentic content pipeline that has run in production for a year

A spaced-repetition study platform for around 13,000 students. I built the LangGraph agent that turns source material into MCQs, flashcards and study notes without an educator writing them by hand, plus the real-time chat layer around it.

~13,000
total registered users(internal)
4,000+
learning items generated in production(internal)
role
AI Engineer
status
live
when
2025 to present
  • LangGraph
  • LangChain
  • RAG
  • Python
  • FastAPI
  • FAISS

Semantic search over an infinite note canvas

Sticky notes on an infinite canvas that find each other. Every note is embedded on write, and a pgvector nearest-neighbour query surfaces related notes without anyone tagging anything. Built and shipped solo.

top 5 of 6
pgvector nearest neighbours, self dropped
text-embedding-3-small
embedding model, one vector per note
role
Solo, design through deploy
status
live
when
2025
  • Next.js 16
  • React 19
  • TypeScript
  • Tailwind v4
  • Supabase
  • Postgres

392 lines, no ML libraries, and a bug that was not where I looked

A feedforward network in one C file with nothing but libc and libm. Forward and backward propagation, He and Xavier initialisation, Box-Muller sampling, binary cross-entropy. The interesting part is not that it trains, it is why it would not.

81.01%
test accuracy
8 to 4 to 1
architecture, seed 42, reproducible
role
Solo
status
live
when
2026
  • C
  • gcc
  • no ML libraries

Where I have worked

  1. Spaced Revision

    ·AI Engineer

    2025 to present · Sikkim, India

    • Built and deployed a LangGraph agentic content-generation system that uses multi-step RAG to produce MCQs, flashcards and study notes from source material, generating 4,000+ items in production over the past year and cutting educator content-preparation time by 60%.
    • Designed and shipped a real-time chat service on Socket.IO and Express supporting 1:1 and group channels, spanning three stores across two services: messages and read state in MongoDB, identity and entitlement in MySQL, and Redis-backed queues driving push fan-out. Room and conversation keys are derived rather than negotiated, so both participants resolve the same identifier independently. In production for 9+ months.
  2. Bytesberry Technologies

    ·Software Development Intern

    2024 to 2025

    • Worked on a headless CMS in React and Node/Express, implementing role-based access control and contributing across the application lifecycle.
    • Worked in an Agile team through the full release cycle, contributing to code reviews and cross-functional technical planning.
  3. Waglogy Tech LLP

    ·Data Analyst and AI Engineer

    2022 to 2025 · Freelance, concurrent

    • Built an LLM-powered automation pipeline that extracted and classified client data, routed customer-support requests, and folded Google search-trend data into SEO work on emerging Sikkim tourism topics.
    • Owned the extraction and routing layer, removing the manual data-entry step from the delivery process entirely.

What I work with

AI / GenAI
LLM applications, RAG pipelines, Agentic systems, OpenAI API, PyTorch, Embeddings
Agent frameworks
LangGraph, LangChain, LangSmith
Backend
Python, FastAPI, Node.js, Express, REST, SQL
Data & retrieval
Vector databases, pgvector, FAISS, ChromaDB, PostgreSQL, MySQL, Redis, MongoDB
Infrastructure
Docker, AWS, GitHub Actions, CI/CD
Frontend
TypeScript, React, Next.js, Tailwind CSS

Get in touch

The fastest way to reach me is email. I read everything and reply to anything that is not a template.