Latest tools · Recent work · About · Contact
My seven newest repositories, ordered newest first. Practical tools for testing agent behavior, validating context, and checking software changes — available on npm or PyPI.
| Project | What it helps you do | Install |
|---|---|---|
| RetrievalCI | Catch retrieval regressions with corpus snapshots, reviewed queries, and evidence-based reports. | PyPI |
| RepoContract | Turn TypeScript architecture rules into executable import-boundary and dependency checks. | npm |
| PatchProof | Check that a regression test fails before a patch and passes after it, with inspectable evidence. | PyPI |
| FixtureForge | Give agents repeatable ticket-service simulations with persistent state and explicit success checks. | npm |
| MemoryLint | Find stale repository-memory claims and propose evidence-backed corrections for review. | PyPI |
| BrowserAssert | Verify browser-action outcomes on an existing Playwright page with evidence-backed postcondition checks. | npm |
| ToolChaos | Inject deterministic tool faults to test how agents recover from failures and missing responses. | npm |
Each repository includes installation steps, examples, and architecture documentation.
More recent projects, ordered by repository creation date, newest first.
| Repository | Focus |
|---|---|
| Agent Systems Lab | Interactive playground for agent-backend concepts, including scheduling, retries, budgets, permissions, and tracing. |
| fabric-agent | Query a Microsoft Fabric semantic model with DAX through an OpenAI-compatible agent and an inspectable console. |
| agent-platform | Relay: a Next.js and FastAPI workspace for agents, scheduled runs, MCP tools, connectors, and team workflows. |
| glossary | Search engineering definitions, production failure modes, and deeper explanations in one field manual. |
| ccmeter | Analyze local Claude Code transcripts with cost estimates, project budgets, exports, and a browser dashboard. |
| promptforge | Refine prompts and explain improvements in browser chat inputs and coding agents using a shared TypeScript engine. |
I engineer intelligent systems end-to-end — from designing multi-agent architectures and fine-tuning LLMs, to shipping the typed APIs, data layers, and frontends that put them in production.
- 🤖 Agentic AI on MCP — custom clients, servers & multi-agent orchestration
- 📚 RAG + fine-tuning across OpenAI, Claude, Gemini & Mistral
- ⚙️ Typed production backends — FastAPI, Pydantic, Postgres, Redis
- 🌐 Full stack delivery — React / Next.js / TypeScript
- 🌱 Currently: Google ADK + function-tool development
class BalajiNaik(Engineer):
def __init__(self):
self.roles = ["Applied AI Engineer", "ML Engineer", "Full Stack Engineer", "Backend Engineer"]
self.ai_stack = ["Agentic AI (MCP)", "RAG Pipelines", "LLM Fine-Tuning", "Multi-Agent Orchestration"]
self.llm_apis = ["OpenAI", "Claude", "Gemini", "Mistral", "OpenRouter"]
self.ml_depth = ["Supervised", "Unsupervised", "Reinforcement Learning", "NLP", "Computer Vision"]
self.backend = ["FastAPI", "Pydantic", "Node.js", "PostgreSQL", "MongoDB", "Redis"]
self.frontend = ["React", "Next.js", "TypeScript", "Tailwind"]
self.now = "Google ADK integration + Function-tool / agent-as-a-tool development"
def philosophy(self) -> str:
return "Models are easy. Systems are hard. I build the systems."
Multi-agent architectures on Model Context Protocol — custom MCP clients & servers, agent-to-agent collaboration, function tools, and agent-as-a-tool patterns. Currently extending with Google ADK. |
Production RAG pipelines with vector databases, hybrid retrieval, and grounded generation. Fine-tuning workflows for domain-specific models across OpenAI, Claude, Gemini & Mistral. |
Typed, scalable APIs with FastAPI + Pydantic, async processing, caching with Redis, and data modeling across PostgreSQL / MongoDB. Containerized with Docker, deployed on AWS. |
Full ML lifecycle — supervised, unsupervised & reinforcement learning with PyTorch / TensorFlow, from feature engineering to training, evaluation, and deployment. |
Complete products with React / Next.js / TypeScript frontends over Node & Python backends — auth, dashboards, real-time features, and clean UX. |
Consistent problem solving on LeetCode — data structures, algorithmic thinking, and complexity-aware engineering carried into every system I design. |
flowchart LR
U[👤 Client / UI<br/>Next.js · React] --> G[🛡️ API Gateway<br/>FastAPI · Pydantic]
G --> O[🧠 Agent Orchestrator<br/>MCP Client]
O --> A1[🤖 Agent: Research<br/>MCP Server]
O --> A2[🤖 Agent: Reasoning<br/>MCP Server]
O --> A3[🛠️ Function Tools<br/>Google ADK]
A1 --> R[(📚 Vector DB<br/>RAG Retrieval)]
A2 --> L[⚡ LLM Layer<br/>OpenAI · Claude · Gemini · Mistral]
A3 --> X[🌐 External APIs<br/>& Services]
R --> L
L --> O
O --> G --> U
Explore earlier projects and engineering work
|
Specialized agents over a shared retrieval layer. Orchestrator routes queries to research/reasoning agents; hybrid retrieval grounds every generation; MCP handles agent-to-agent context passing.
|
Dataset → training → evaluation, automated. End-to-end pipeline for adapting foundation models to domain-specific tasks, with reproducible configs and evaluation gates before release.
|
|
One typed interface, many LLM providers. FastAPI gateway unifying OpenAI, Claude, Gemini & Mistral behind consistent schemas — with routing, fallback, and validation via Pydantic.
|
Documents in, structured intelligence out. Claude- and Gemini-powered extraction, classification, and analysis over unstructured documents.
|
|
Reinforcement learning from scratch. Reward-driven agent that learns Snake through deep Q-learning in an interactive PyTorch + Pygame environment.
|
Full-stack, end to end. Ordering platform with auth, cart & category flows, and a complete admin dashboard for catalog management.
|
| 🎯 | What you get |
|---|---|
| End-to-end ownership | I take features from architecture diagram → model → API → UI → deploy. No hand-offs needed. |
| AI that ships | Not notebook demos — production RAG, agent systems, and fine-tuned models behind typed, tested APIs. |
| Engineering discipline | Pydantic-validated boundaries, Dockerized services, DSA-grounded problem solving (active on LeetCode). |
| Current stack, not last year's | Already building on MCP and Google ADK — the newest agent tooling in production patterns. |
Open to roles and collaborations in Applied AI, ML Engineering, Backend & Full Stack — especially anything involving agents.
