_{State → Route → Execute → Validate → Observe}

   

⚡ Status: Actively interviewing for AI Engineer roles (India / Remote)

---

🧩 Core Capabilities

Multi-Agent Systems Stateful Workflows Deterministic Routing Hybrid RAG Tool Execution HITL Cost-Aware AI Evaluation & Guardrails

---

⚡ Engineering Philosophy

"The architecture meeting you skip always becomes the non-deterministic bug you can't trace in production."

I build reliable LLM applications, multi-agent orchestrations, and hybrid retrieval systems. My approach prioritizes explicit state machines, strict context boundaries, cost-guardrail routing, and deterministic evaluation layers before relying on raw LLM inference.

---

🚀 Active Build

🧭 Waypoint — Multi-Agent AI Trip Planner

Every great journey starts with a Waypoint. A user describes a trip through a structured form; a graph of LangGraph agents plans transport, lodging, activities, and budget; the system produces a day-by-day itinerary; and the user refines it through a chat-based review loop before finalizing.

User flow: Form fill → Trip session starts → Planning runs → Human review (chat loop) → Final itinerary

• 5-Agent Graph: Concierge normalizes raw form input → Planner decides search priority → Budget sums cost against stated budget → Itinerary composes the day-by-day plan → Critic interprets edit requests during review. Human-in-the-loop gate at human_review node.
• Write-Safe State: TypedDict TripState — single source of truth with separate top-level keys per tool result (not a grouped dict) for conflict-free parallel fan-out. Postgres checkpointer (langgraph-checkpoint-postgres) persists paused graph state so a session survives refresh/restart.
• 7 Mechanical Tool Nodes: search_flights, search_trains, search_buses, search_cars, search_hotels, search_activities, get_weather — no LLM calls, staged dummy → free API (Tavily) → MCP.
• First Framework Project: LangGraph + LangChain orchestration instead of hand-rolled agents — deliberate framework adoption with full understanding of the graph execution model underneath.
• Stack: Python • LangGraph • LangChain • FastAPI • Postgres • SQLAlchemy • Alembic • Pydantic • Tenacity • Groq (gpt-oss-120b) • OpenAI (fallback) • Tavily • MCP

---

🛠️ Featured Systems & Architectures

💸 FinOS — AI Finance Platform Full-stack financial OS combining an interactive analytics dashboard with a conversational agent engine — both writing to the same SQLite store. Zero-Cost Hybrid Router: Resolves ~60% of deterministic queries via pattern matching at $0 LLM cost; streams fallbacks via Groq + SSE. Predictive Analytics: Financial health scoring, spending trajectory forecasts, budget threshold alerts. Hardened Auth: bcrypt with session tokens and CLI-based recovery. Python • FastAPI • SQLModel • Groq • SSE • SQLite	🧠 hArI v2 — Document Intelligence High-precision RAG engine rebuilt from a single-store ChromaDB prototype into a hybrid-search, multi-user system. Hybrid Retrieval: Dense pgvector similarity coupled with tsvector full-text search via a custom Supabase RPC. Sandboxed Analytics: DuckDB SQL engine replacing pandas.exec() for safe multi-CSV queries. Multi-User: Supabase Auth, per-user chat persistence, citation UI with telemetry feedback. Python • Supabase • pgvector • DuckDB
🤖 AI Agent Engine Modular 4-layer query routing and execution system designed to eliminate unconstrained agent loops. Deterministic Pre-Filtering: ChromaDB + SentenceTransformer semantic layer resolves ~80% of repeat queries locally at ~30ms. Guardrail Architecture: Planner → Validator → Executor → Responder with hard session token and quota cutoffs. Cost Efficiency: Per-session cost ~$0.0005. Python • Gemini API • ChromaDB • SentenceTransformer	🔬 Research Pipeline 6-agent supervisor workflow orchestrating deep, multi-hop web research with zero context degradation. Stateful Segregation: Supervisor → Search → Scraper → Summarizer → Critic → Synthesizer. Model Routing: Groq (llama-3.3-70b) primary Gemini (gemini-2.0-flash) fallback, complexity-based. HITL Control Gates: Human-in-the-Loop review checkposts with multi-provider fallback layers. Python • Groq • Gemini API • Rich CLI
🎯 NextSteps — Career Gap Engine Actionable talent intelligence pipeline translating JD criteria into verified technical roadmaps. Semantic Parsing: Unstructured resume extraction matched against live JD requirements or URLs. Targeted Synthesis: Maps identified skill discrepancies directly to granular learning modules. Clean Pipeline: Parse → analyze → output — no hardcoded resume/JD assumptions. Python • FastAPI • Groq • Tavily

---

🗃️ Tooling & Protocol Implementations

Repository	Focus & Architecture	Primary Tech
🔌 DevMind — MCP Server	Secure Model Context Protocol tool server exposing 6 developer tools — file ops, Python execution, JSON utilities — with human-gated execution boundaries.	MCP SDK • Python • tiktoken
🗄️ LangGraph Parallel SQL Runner	Graph-governed parallel SQL executor via LangGraph's Send API, with automated syntax validation and pre-execution human review.	LangGraph • Groq • SQLite

---

💻 Tech Stack

🤖 AI & Agent Systems

🗄️ Data & Vector Storage

🐍 Languages & Backend

🛠️ Tools & Interfaces

---

📈 Activity & Metrics

---

Interested in collaborating or discussing AI Engineer opportunities?