I build production AI systems where something real is at stake. Tight budgets, no internet, live phone calls — constraints that make the standard solution impossible are where I do my best work.
| 94% AI cost reduction | Rebuilt inference pipeline from scratch, Rs. 350 to Rs. 20 per unit in production |
| Rs. 30L+ saved annually | Multi-agent automation pipelines replacing entirely manual workflows |
| Voice AI under 500ms | Real-time AI agent on live phone calls, custom Cerebras CS-2 inference architecture |
| Full LLM offline in 1GB | Quantised LLM on-device on budget Android hardware, zero cloud dependency |
🎙️ Aurora — Real-Time Voice AI Agent
Standard LLM APIs take 3 to 5 seconds. That makes natural phone conversation impossible. Built a custom 3-stage inference architecture — hitting under 500ms end to end. Live in production at 99.7% success rate.
🎨 agentdraw-canvas — Published npm Package
Canvas engine that gives AI agents a programmatic drawing hand. Every shape gets a UUID so agents create, update, animate, or delete elements across sessions and multi-step operations. Plugin architecture across 13 independent ES modules. 111 weekly downloads at peak.
I write about productionizing AI, solving edge cases, and moving beyond basic GenAI wrappers on Medium & Towards AI.
- 🧠 This Is Why Your Model Hallucinates (And You Blame the Wrong Thing) — Techniques for arithmetic validation and structural grounding.
- ⚙️ How to Think Like a Prompt Engineer (Not Just Write Better Prompts) — Replacing AI judgment with algorithmic protocols in messy real-world extraction.
- 📖 What Is Machine Learning Really? A Human Guide to the Most Powerful Idea in Tech — Explaining ML optimization vs memorization.
- 📖 Authored a chapter in a published book on Artificial Intelligence
- 📜 Filed 6 copyrights with 1 patent application in process
- 🏆 Won national-level project competitions 3 times
- 👥 Mentored 20+ students on AI implementation
I'm currently open to new opportunities and consulting. If you are dealing with heavy optimization problems, architecture bottlenecks, or need custom multi-agent rollouts, let's talk.
The problems I enjoy most are the ones where the standard solution doesn't exist yet.