I’m an AI engineer based in Ahmedabad, India. In under two years I’ve shipped seven end-to-end AI systems into production — a permission-aware agentic workflow, RAG chatbots serving thousands of clients, multi-model LLM pipelines, and on-device ML models running in 2 KB of RAM.
The thread connecting all of it: I prefer systems that real users depend on daily, not demos. That means caring about permission systems, latency budgets, on-prem constraints, evaluation that survives contact with production data, and benchmarks rather than vibes for model selection.
Before that — research at IIT Delhi, embedding work with Prof. Mayank Singh at IIT Gandhinagar, and two TinyML hardware projects that reached national finals (Samsung Top 10 of 70,000; Smart India Hackathon Top 5 of 44,000; Ingenium 2nd of 200+).
SHIPPED THREE PRODUCTION AI SYSTEMS IN UNDER A YEAR.
A LangGraph agent over the whole platform: intent detection routes each turn to RAG search, Postgres text-to-SQL, or one of 25+ tool APIs (filtered data fetches and CRUD actions like create-user). A permission layer gates every action first — 250+ permissions, 750+ courses, no caching possible.
PDF-to-form converter (~20s per form) plus a conversational builder. Three LLMs — GPT-4o, Qwen3-Coder-480B, GPT-4o-mini — bridged by a Pydantic-validated intermediate schema.
10,000+ clips embedded with Cohere, retrieved against Milvus, re-ranked, ordered and narrated by Gemini 2.5 Pro, voiced by ElevenLabs, rendered by FFmpeg on AWS Lambda. 3-phase human-in-the-loop.
INHERITED A SLOW STACK. LEFT IT AT 95% MRR.
Decoupled Llama-2-7b into a vLLM-served API. Benchmarked vector DBs on production-scale data — Milvus 95% / Qdrant 85% / FAISS 45% MRR — and led the migration. Evaluated 6+ embedding models, landed on BGE-M3 + reranker with custom query preprocessing.
Aadhaar, PAN, GST — data could not leave the client’s 250 GB server. Hybrid OCR (DocLink + DocTR) merged by Nemotron-70b hosted on-prem. Entity extraction layered on top.
NL → SQL over a live SAP DB. 110+ table chunks (DDL + sample rows) indexed in Qdrant, LLM query expansion, self-refinement for missing JOIN tables, exactly 2 auto-generated charts. ~89% end-to-end success.
A WEARABLE SYSTEM. HONEST EVALUATION.
ESP32 + MPU6050 wristband at 60 Hz over Wi-Fi. Collected 20 participants × 62 classes × 10 repetitions = 12,400 samples. Swept Dense / 1D-CNN / LSTM / BiLSTM / hybrids across 1,000+ configs. CNN-BiLSTM final.
Wearable air-writing recognition system. Hardware, data, models, paper. See Role 03 above for the full case.
Worked with Prof. Mayank Singh on embedding models and retrieval systems. Foundations for the later production RAG work were laid here.
Co-authored a research paper on infant cry classification with embedded sound recognition.