Cityflo
Software Development EngineerJul 2025 – Present · Mumbai- Pitched, prototyped in a 6-hour hackathon, and shipped Komo AI to production — a multi-agent service platform now used company-wide by 80+ people across engineering, ops, support, and founders — cutting investigation time from 30–60 minutes to 2–3 minutes across 30–40 daily investigations.
- Designed the agent framework itself: spec-driven agent definitions, per-agent MCP tool servers, USD-budget-bound execution, and bidirectional agent handoff. Shipped two agents — Tech Mazdoor (17 tools) and CX-Pulse (11 tools) — integrating Aurora, ClickHouse, GitHub, Datadog, Sentry, Airflow, Celery, Linear, and CleverTap.
- Built core components of SageAI, Cityflo's AI customer support backend in Go — 36K chat contacts/week with an 80% AI deflection rate, cutting human-handled contacts per 1,000 rides by 56%. Owned the LLM integration, AI→human handoff orchestration, and ticket sync.
- Replaced SageAI's Flowise layer with an in-house agent service — FastAPI + PydanticAI agents on a conversation graph exposing 20+ tools across ride, ticket, payment, and driver domains, a fallback-wrapped LLM cascade via LiteLLM, and a realtime voice agent (Gemini Live over LiveKit WebRTC, call recordings to S3).
- Engineered SageAI's per-room concurrency layer for stateful AI conversations across horizontally-scaled ECS: goroutine-per-room with buffered channels, Redis SetNX locks for exactly-one agent spawn, and Redis Pub/Sub for cross-instance command delivery.
- Own Cityflo's new B2C subsegment — Luxe premium cabs and Bus Rentals — end to end: zone-based dynamic pricing with direct ops control, an operator dashboard that lifted ops efficiency 40%, API latency cut from 1s to 200ms, and a direct customer booking flow now going live.