Agentic AI Platform

Overview

Built the Agentic AI runtime for a global consumer AI chat platform with 900K+ paid subscribers and $1M+ daily net revenue. The goal was to evolve the product from a chat wrapper into a true agentic platform — multi-step reasoning, real-time tool use, and consent-gated actions — shipped as a zero-downtime additive layer on a live production system.

Achievements

• Shipped a production-ready agentic MVP across 5 phases with zero downtime, enabling the product to compete directly with ChatGPT and Manus. • Headed the Connectors department (team of 11) and designed its entire operational process from scratch — an 8-stage Auth → Backend → Frontend → QA → Release pipeline that ships 5–8 connectors to production per day without interrupting ongoing development. Embedded Claude-powered AI agents at every stage so engineers supervise and approve rather than execute, enabling the platform to offer 4,000 integrations where users read emails, search docs, create tasks, and send messages — all from a single AI chat. • Drove P0 cost optimisation on daily inference spend — model routing, context compaction, prompt A/B testing — protecting margins while guarding output quality via LLM-as-judge eval gating.

Responsibilities

• Architected the full agentic runtime: context assembly layer, multi-turn agent loop with live tool calls, streaming responses, error recovery, and token budget management across 12 global regions.
• Designed the MCP-first connector architecture with zero-central-edits isolation — each connector is a self-contained folder discovered by codegen, so parallel work across dozens never causes merge conflicts.
• Built the LLM evaluation and prompt-optimisation framework: 5-dimension LLM-as-judge scoring, failure-driven candidate generation, and automated weekly prompt improvement via Langfuse.

Technologies Used