Work
Over two decades, I have architected and delivered global-scale platforms used by millions — and recently built a production-grade agentic AI platform from first principles. Below are selected highlights from my portfolio — starting with my recent AI work
AI Portfolio
Production-grade agentic AI systems built from scratch — designed, architected, and measured. All repos at github.com/manisundaram.
agentic-ai-platform — Flagship
LangGraph · LlamaIndex · MCP · LangSmith · RAGAS
A production-ready agentic AI platform built end-to-end as sole architect and engineer. Demonstrates enterprise patterns including stateful agent orchestration, retrieval grounding, tool execution, observability, and evaluation.
What it does:
- Stateful multi-agent orchestration with LangGraph — typed state, conditional cycles, and Human-in-the-Loop (HITL) interrupt/approval flows
- RAG pipeline via LlamaIndex VectorStoreIndex + ChromaDB with hybrid search
- MCP (Model Context Protocol) server exposing standardized tools to the agent
- Full observability via LangSmith tracing; evaluated with RAGAS
Measured results:
| Metric | Score |
|---|---|
| Agent Task Completion | 1.00 |
| Tool Call Accuracy | 1.00 |
| RAG Faithfulness | 1.00 |
| Hallucination Rate | 0.10 |
| Overall RAGAS Score | 0.613 |
Stack: Python, LangGraph, LlamaIndex, FastAPI, ChromaDB, Docker, GitHub Actions, Railway, LangSmith, RAGAS
agents-api — Custom ReAct Multi-Agent System
Python · FastAPI · OpenAI · Anthropic · Gemini · Prometheus
Built a ReAct agent loop from primitives — deliberately without a framework — to understand the mechanics before abstracting them. Implements a Planner → Worker → Reviewer multi-agent pattern with production-grade engineering.
What it does:
- Custom ReAct loop with model routing, semantic caching, and guardrails
- Multi-provider factory pattern supporting OpenAI, Anthropic Claude, and Google Gemini
- PII masking, Prometheus metrics, and structured logging throughout
Stack: Python, FastAPI, Docker, GitHub Actions, Prometheus
ai-service-kit — Shared Library
Python · pytest · AWS · Azure · GCP · Datadog
A production-grade shared library used across all sibling services. Built for reliability, testability, and multi-cloud portability.
What it does:
- 102 passing tests with deterministic mock providers (SHA-256 seeding)
- 2-level provider fallback for resilience
- Cloud logging abstraction supporting AWS, Azure, GCP, and Datadog
- Companion template repo provides operational endpoints (health, diagnostics, metrics, ping), CORS, and request logging out of the box — so new services start production-ready
Stack: Python, pytest, Docker
GitHub - AI Service Kit | GitHub - AI Service Template
semantic-search-api — Semantic Search Service
Python · ChromaDB · FastAPI · Embeddings
A standalone semantic search service with hybrid search, multi-provider embedding support, and a health dashboard.
What it does:
- Embeddings with multi-provider support (OpenAI, BGE, Instructor)
- ChromaDB vector store with hybrid search and chunking strategies
- Health dashboard for monitoring search quality
Stack: Python, FastAPI, ChromaDB, Docker
rag-api — RAG Pipeline from Primitives
Python · FastAPI · ChromaDB
Built a complete RAG pipeline from scratch before adopting LlamaIndex — deliberately understanding document loading, chunking, embedding, and retrieval at the primitive level before abstracting.
What it does:
- Document loader, chunker, embedder, and retriever built from first principles
- Informed the architecture decisions made in agentic-ai-platform
Stack: Python, FastAPI, ChromaDB, Docker
llm-chat-api — Foundation Chat Service
Python · FastAPI · Multi-provider
The foundational chat service and provider abstraction baseline that all other services build on.
Stack: Python, FastAPI, Docker
Selected Product Portfolio
- WHO Learning Portal: Architected a global health education platform serving 50,000+ users worldwide.
- iSafe: Led the development of an AI/IoT-powered student safety system featuring facial recognition and RFID tracking, deployed in 50+ schools.
- Atrial Rhythm: Designed a mobile healthcare platform for provider-patient record management and scheduling.
- Enterprise Cloud Migration: Directed a massive database migration for USPS with near-zero downtime under strict SLAs.
Intellectual Property
I hold two U.S. Patents focusing on the future of data and storage:
- Virtual Private Cloud Storage: Combined cloud-based and physical storage environments (US 14/144,970).
- Contextual Information Presentation: Organizing analytically relevant info for high-efficiency display (US 13/714,329).
Leadership & Entrepreneurship
As the Founder of Inovark Technologies, I lead a global engineering firm delivering SaaS and AI-enabled platforms across the U.S., Europe, and APAC. I specialize in scaling engineering teams from the ground up and driving 30% gains in delivery velocity through agile and AI-assisted workflows.
For a full chronological history, please visit my LinkedIn Profile.