Sumit Singh — AI Engineer

01 — Voice + RAG

AI voice assistants that answer from a client's own knowledge base — speech in, retrieval against their documents, speech out.

PythonRealtime voiceRAGAWS

View project

Problem: Support teams answering the same questions by phone, from documents no assistant could read.
Built: Streaming speech pipeline over a retrieval layer, with the knowledge base editable by the support team — no deploy to change an answer.
Outcome: Assistants that stay current because the people who know the answers own them.

02 — Retrieval infrastructure

Drop in a PDF or DOCX and get a working retrieval system back, immediately.

PythonNext.jsVector search

View project

Problem: Every project that needed retrieval started by rebuilding the same parsing and chunking pipeline.
Built: Upload to queryable in one step: extraction, chunking, embedding and a query endpoint behind a single interface.
Outcome: Retrieval became a starting point instead of a week of setup.

03 — Model evaluation

A benchmark harness for comparing language models on what actually decides the choice.

TypeScriptReactMulti-provider APIs

View project

Problem: Model choice was being made on published benchmarks that say nothing about cost on your own workload.
Built: One harness that runs the same task across providers and reports token usage, price per task and output quality side by side.
Outcome: A model decision defensible with your own numbers.

I build AI systems that ship to real users.