SUMANTH KUMAR M · AI / AGENT ENGINEER · 8 YEARS
I design, ship, and run multi-agent LLM systems in production — autonomous software that runs 24/7, not prototypes. Open-source: a multi-agent code reviewer with a measured eval benchmark (Quorum), an AI operating system with seven specialist agents (Aria), and a fully on-device private AI (Airgap). The delivery proof behind all of it: 8 years shipping production software — 15+ mobile apps at 100K+ installs, 99.5% crash-free.
Three open-source agent systems with live demos, a production multi-agent ops platform, and 15+ shipped mobile apps as the delivery track record behind it.
A real LLM runs fully on the device and answers questions about your own documents via on-device retrieval — with citations, claim-level self-verification, and an honest confidence read (it tells you when it isn't sure). Offline after a one-time model download; no cloud, no API key. Imports PDFs, voice in & out, swappable 0.5B/1.5B model. Flutter + native (MediaPipe LiteRT / llama.cpp) · Apache-2.0 · tested + CI.


A dispatcher routes a diff to specialist LLM agents — security, performance, correctness, tests, style — in parallel; a synthesizer consolidates them into one verdict. Provider-agnostic (Claude / OpenAI / Ollama) with native structured output, a CI-gated detection-eval benchmark (recall / false-alarm scored), config-driven custom agents, and zero runtime dependencies. The demo runs the whole engine in your browser via WebAssembly.
A macOS-style web desktop with a multi-agent runtime underneath. An orchestrator decomposes a goal and routes it to seven specialist agents that search the live web with citations, write and execute real code (Python on WebAssembly, sandboxed JavaScript), generate images, and review one another — streamed into a live mission-control graph. Brain-agnostic by design: the same agents run on a deterministic simulated engine, your own API key, or a real LLM running entirely in your browser via WebGPU — no server, no key. Next.js · TypeScript · unit-tested · CI.
An LLM dispatcher routing tasks to specialist agents (triage, content, code execution, strategy), with a realtime task board (REST + SSE), per-task model routing, retries, heartbeats, idempotency, and agents that self-report via API.
15+ production apps (Flutter / Android / iOS) with Firebase, CI/CD, 15-language localization, and on-device + cloud AI — 100K+ combined installs at 99.5% crash-free.
Deep in LLM / agent engineering, with 8 years of production software delivery behind it.
Multi-agent orchestration (custom dispatch/synthesis) · on-device / local LLMs · RAG & vector search · tool use & function calling · MCP · structured output · prompt engineering & caching · evals (golden-dataset, LLM-as-judge) · Claude · OpenAI · Ollama
Flask · REST · Server-Sent Events · SQLite · Firebase · CI/CD · Linux · systemd · cron · Cloudflare Workers
Native Android (Kotlin · Java · Jetpack Compose · MVVM · Coroutines) · Flutter · React Native · iOS · Firebase · Play Store release · 15-language localization
Python · Kotlin · Java · Dart · Swift · JavaScript / TypeScript
8 years of software engineering — production LLM/agent integration on top of a deep mobile engineering base (native Android, React Native, Flutter).
AI · Mobile LLM/agent integration on production mobile apps, plus React Native and native Android development and Android device-session tooling.
Mobile Built and shipped cross-platform and native mobile apps — React Native, Flutter, and native Android (Kotlin, Jetpack Compose, MVVM, Coroutines) — with Firebase, CI/CD, and analytics across client products.
Android Native Android development for a logistics platform — Kotlin, MVVM, and Coroutines — building and scaling features for delivery operations at scale.
Mobile Native Android and cross-platform (React Native / Flutter) feature development across client products, from build to Play Store release.
Android · Voice Android app development with voice and speech features — building mobile experiences around speech recognition and audio.