ai engineer building llm-powered products — production rag pipelines with real evals, multi-provider llm tooling, and the python apis behind them. currently software engineer @ metquay. india · open to remote.
[w] prayagtushar.xyz · resume [x] x / twitter [l] linkedin [m] t.prayag.eng@gmail.com
[s] skills
- ai / llm — rag, hybrid search + reranking, embeddings, llm evals (llm-as-judge), langchain, langfuse
- providers — openai, anthropic, gemini
- backend — python, fastapi, nestjs, node.js
- languages — typescript, javascript, java, c++
- data / infra — postgresql, pgvector, pinecone, redis, docker, gcp cloud run, aws
- frontend — next.js, react, tailwind
[p] projects
-
isra — hybrid-retrieval rag over the indian startup ecosystem. pgvector + postgres full-text fused via reciprocal rank fusion, bge cross-encoder reranking (no langchain), llm-as-judge evals (mrr 0.688 → 0.750), streaming sse chat with grounded citations, langfuse tracing. live · repo
-
multi-llm client — unified async python client for openai, anthropic & gemini. one interface for messages, streaming, usage, and errors; ships as library, cli, repl, and fastapi service. pydantic v2, mypy-strict, 40 tests. repo
-
readora — chat with your pdf. rag over gemini embeddings + pinecone, next.js 15, clerk, neon. live · repo
-
lumosai — gemini chat assistant with persistent threads. mongodb + clerk. live
-
kestra todoist plugin — 1,300+ loc contribution merged into the open-source workflow orchestrator. kestra
[b] writing
[c] say hi
open to interesting problems, especially around ai infra and llm tooling. let's build something cool.