I build Python and TypeScript backends and the tooling around them, and I try to measure whether what I ship works.
DiffContext (https://pypi.org/project/diffcontext/) is a context-selection library for code LLMs. It cuts the context a model reads by a median 89% across 9 repos and 701 commits, and raised pass@1 from 5.5% to 25.8% on 128 tasks. It ships as a CLI and MCP server with 358 tests in CI, and I published a self-audit that retracted three of my own earlier claims.
Open source, every fix with tests: merged in Comet Opik (streaming wrappers were swallowing exceptions) and facebook/astryx (focus-trap bug). In review: apache/superset (approved), kubeflow/mcp-server (FastMCP 4 upgrade), bitcoindevkit/bdk_wallet (Rust, relative locktimes), and a Java/ClickHouse change in Opik. I also review PostgreSQL patches on CommitFest.
Before that: React front ends on Django REST and FastAPI back ends over PostgreSQL (cut API response times 35%), and FastAPI model serving at 150 req/s. Now building Python environments that AI agents are evaluated in.
M.Tech in AI & Data Science at IIT Patna, part-time.
Looking for backend, AI/LLM or full-stack roles. Can start immediately.
Remote: Yes
Willing to relocate: Yes (would need visa sponsorship outside India)
Technologies: Python, TypeScript, SQL, FastAPI, Django REST Framework, Node.js, React, Next.js, PostgreSQL, SQLite, MongoDB, Cloudflare Workers, Docker, Linux. LLM systems: agents, tool calling, MCP, RAG, embeddings, BM25 retrieval, evals. Read and patched Java, Rust and C in open-source work.
Resume/CV: https://drive.google.com/file/d/1MK3XqvWlbDTq2l1MVm7yXGKiMfS...
GitHub: https://github.com/trakshan-mishra
Email: [email protected]
I build Python and TypeScript backends and the tooling around them, and I try to measure whether what I ship works.
DiffContext (https://pypi.org/project/diffcontext/) is a context-selection library for code LLMs. It cuts the context a model reads by a median 89% across 9 repos and 701 commits, and raised pass@1 from 5.5% to 25.8% on 128 tasks. It ships as a CLI and MCP server with 358 tests in CI, and I published a self-audit that retracted three of my own earlier claims.
Open source, every fix with tests: merged in Comet Opik (streaming wrappers were swallowing exceptions) and facebook/astryx (focus-trap bug). In review: apache/superset (approved), kubeflow/mcp-server (FastMCP 4 upgrade), bitcoindevkit/bdk_wallet (Rust, relative locktimes), and a Java/ClickHouse change in Opik. I also review PostgreSQL patches on CommitFest.
Before that: React front ends on Django REST and FastAPI back ends over PostgreSQL (cut API response times 35%), and FastAPI model serving at 150 req/s. Now building Python environments that AI agents are evaluated in.
M.Tech in AI & Data Science at IIT Patna, part-time.
Looking for backend, AI/LLM or full-stack roles. Can start immediately.