I take AI from prototype to production, with a focus on systems that are reliable, observable, and built to last.
CLI tool that automates the full print-on-demand pipeline — from AI image generation to SEO-optimized listings and scheduled RedBubble uploads. Built for freelancers who want to ship designs without the manual overhead.
A full-featured Adobe CC add-on for advanced video captioning, built with React and TypeScript inside Adobe's Document Sandbox Runtime. Supports customizable caption styles, local storage, and a responsive UI across Adobe applications.
Full-stack app that lets users describe a mood in plain English and get a matched Spotify playlist back. A React frontend talks to a Python backend that maps natural language to Spotify's audio feature API — because keyword search doesn't understand feelings.
Streamlit-based RAG app that lets users upload documents and query them in natural language. Uses ChromaDB for vector storage and Gemini as the reasoning layer — making document search feel like talking to someone who actually read everything.
Every system I build follows a composable, evaluation-first approach. Agents should be testable in isolation, observable in production, and replaceable without rewriting the orchestration layer. I optimize for debuggability over cleverness, and measure everything.
A deep dive into HNSW parameter tuning, index sharding, and pre-filtering strategies that brought p99 latency under 200ms for 10M+ documents.
Practical strategies for chunking, summarization chains, and dynamic context assembly when your retrieval returns more than the model can handle.
Building automated evaluation pipelines using LLM-as-judge, reference-free metrics, and regression testing for continuous deployment.
Have a project in mind? I work best when the problem is hard and the stakes are real.
© 2026 Upendra Pant. All rights reserved.