AI engineer building RAG chatbots and AI-driven workflow automation. Five years of production Java/Spring backend at Walmart, Oracle, and VMware before that.
I build retrieval systems that admit when they don't know. My document Q&A pipeline: Docling for structure-aware parsing, one batched embeddings call instead of 60 round trips, TimescaleDB + pgvectorscale instead of a hosted vector DB (~75% cheaper), Celery/Redis for async ingestion returning an immediate job ID, and an explicit enough_context flag via Instructor + Pydantic so it returns "I don't know" instead of a confident guess. Langfuse tracing on every token and dollar.
Also built an event-driven agentic platform that classifies inbound email, routes it, acts on it, and escalates to a human when confidence is low — multi-provider (OpenAI/Claude/Gemini/Bedrock), swappable in one line via PydanticAI.
Open to contract or full-time. US Green Card holder — no sponsorship needed. Most useful to teams sitting on a pile of documents who need something reliable on top of them. I'll also tell you when a scheduled Python script beats an LLM — I've done that on a paid engagement and saved the client money.
Gayithri2026 · · focus · HN ↗
Location: San Jose, CA (SF Bay Area)
Remote: Yes
Willing to relocate: No
Email: gayithri@pojoai.com
Technologies: Python, FastAPI, PostgreSQL/pgvector, RAG, LangChain, PydanticAI, Celery, Redis, Docker, AWS, Langfuse, Java, Spring Boot Résumé/CV: <a href="https://github.com/Gayithri606/resume/blob/main/GayithriPonnapalli-Resume.pdf" rel="nofollow">https://github.com/Gayithri606/resume/blob/main/GayithriPonn...
AI engineer building RAG chatbots and AI-driven workflow automation. Five years of production Java/Spring backend at Walmart, Oracle, and VMware before that.
I build retrieval systems that admit when they don't know. My document Q&A pipeline: Docling for structure-aware parsing, one batched embeddings call instead of 60 round trips, TimescaleDB + pgvectorscale instead of a hosted vector DB (~75% cheaper), Celery/Redis for async ingestion returning an immediate job ID, and an explicit enough_context flag via Instructor + Pydantic so it returns "I don't know" instead of a confident guess. Langfuse tracing on every token and dollar.
Also built an event-driven agentic platform that classifies inbound email, routes it, acts on it, and escalates to a human when confidence is low — multi-provider (OpenAI/Claude/Gemini/Bedrock), swappable in one line via PydanticAI.
Write-up: <a href="https://gayithriponnapalli.com/blog/2026/04/14/how-i-built-a-production-ready-ai-document-qa-system--and-what-makes-it-different/" rel="nofollow">https://gayithriponnapalli.com/blog/2026/04/14/how-i-built-a... Code: <a href="https://github.com/Gayithri606" rel="nofollow">https://github.com/Gayithri606
Open to contract or full-time. US Green Card holder — no sponsorship needed. Most useful to teams sitting on a pile of documents who need something reliable on top of them. I'll also tell you when a scheduled Python script beats an LLM — I've done that on a paid engagement and saved the client money.