Patel Tirth · founding ai engineer @ magi · bangalore
I'm Patel Tirth (Tirth Patel on most search results), founding AI engineer at Magi in Bangalore. I build LLM agents, retrieval pipelines and, lately, a voice agent that scores its own calls. Weekend goal: weirder projects. Below is a changelog of everything I shipped, broke and posted, colour-coded the way I write my notes.
me, allegedlyA support-desk voice agent for small businesses, built on Pipecat. It answers from the business's own documents, books slots and checks order status. Local Kokoro TTS, API models for Hinglish speech recognition and the LLM.
Every call is graded automatically: latency per stage (p50 / p95), entity accuracy, groundedness. CI fails on regression.
Any marketing asset, PDF or live URL, checked against the brand's rules. Violations stream over WebSockets and fixes are applied as agentic patches. Featured in Magi's "100x Marketer" series.
Rebuilt my family's tile-catalogue app in Expo so customers can check live stock for any tile. Live on the Play Store. Lesson from the shop floor: tiles are a touch-and-feel purchase, so the app supports the visit rather than replacing it.
Five years of gaming laptop. It survived a motherboard replacement earlier in the year and did not survive me reseating its display cable with the power on. It is now a very expensive power LED. Post-mortem is on X.
One company URL in, a structured brand identity out, via LLM extraction. Parallelised the slow stages for a 46% latency cut. Same quarter: Anthropic's contextual retrieval in the RAG pipeline, and led the Dify to self-hosted Qdrant migration.
The final-year project that started all of this: a RAG pipeline that turns plain-English questions into SQL and blends in unstructured data such as user reviews. VIT filed it as "AI-Powered Business Insights Assistant Using RAG", application 202541118370. A simple system by today's standards. Also the reason I have this career.
First engineer on the AI side. Agents on Agno, RAG, memory and the feedback loop for the platform's Copilot. Remote from Bangalore, one day a week at WeWork.
AI engineer intern at Terra. LangGraph workflows that translate natural language into JQL for Mixpanel data, a 33% token reduction through context optimisation, FastAPI + WebSockets backend with a React front end. Owned the GCP deployment end to end.
Got into AI through a side project rather than campus placements. Also logged roughly 1,200 hours in Valorant as Omen along the way. Not on the resume.
Multiplayer chess over WebSockets with Next.js and server-side rendering. Started as a two-person project; my partner dropped out in the second week and I finished it solo. I don't really play chess. I wanted to finish what I started.
Unicorniz (rebuilt the company site in React and Node), Valsco (front end, plus weekly workshops mentoring junior developers), Kairali Ceramics (first version of the e-commerce app, Play Store beta). The GitHub archive from this era includes a meme generator, a Steam page clone and a chat app.
Code Bullet / Michael Reeves energy: a small, strange build with a reveal at the end. Ideas welcome.
pateltirth.tech@gmail.com
github.com/p-tirth · x.com/_P_tirth_
linkedin.com/in/p-tirth-