AI Systems & Autonomous Agents Engineer building stateful agent workflows, RAG systems, and tool-execution pipelines — architecture first, inference second.
I don't start AI projects by choosing an LLM framework. I start by defining state, execution boundaries, deterministic paths, failure modes, and evaluation. Then I decide where an LLM actually belongs.
I spent 9 months as a full-stack developer at Symbiosis Systems, shipping 3 enterprise apps with ASP.NET and SQL Server. That experience shaped how I think about persistence, APIs, validation, and software that has to survive real usage.
Now I apply that discipline to multi-agent systems, RAG pipelines, tool execution, and cost-aware LLM workflows in Python — using frameworks as implementation tools, not as architecture.
A small portfolio of systems built around explicit state, bounded execution, deterministic routing, retrieval, validation, persistence, and human control.
Multi-agent AI trip planner. You describe 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 you refine it through a chat-based review loop before finalizing.
| Spec | Detail |
|---|---|
| Agents | Concierge to Planner to Budget to Itinerary to Critic. 5 LangGraph agents with human-in-the-loop review |
| Tools | 7 mechanical tool nodes: flights, trains, buses, cars, hotels, activities, weather. Staged dummy to free API to MCP |
| State | TypedDict TripState, separate keys per tool result for write-safe parallel fan-out. Postgres checkpointer survives refresh and restart |
| Stack | Python, LangGraph, FastAPI, Postgres, SQLAlchemy, Pydantic, Groq, Tavily |
| github.com/harshbhanushali26/waypoint | |
RAG document intelligence. Upload PDFs or CSVs, ask questions, get cited answers. Rebuilt from a single-store ChromaDB prototype into a hybrid-search, multi-user system.
pandas.exec() after I realized untrusted CSV queries could execute arbitrary Python.
| Spec | Detail |
|---|---|
| Retrieval | Hybrid: pgvector similarity + tsvector full-text via a custom Supabase RPC |
| CSV queries | DuckDB SQL engine replacing pandas.exec() for safe multi-CSV queries |
| Users | Supabase Auth, per-user chat persistence, citation UI with telemetry feedback |
| Live demo | harichat.streamlit.app (sleeps after 12h inactivity) |
| Stack | Python, Streamlit, Supabase, pgvector, DuckDB, Groq, LangChain, SentenceTransformer |
| github.com/harshbhanushali26/hArI | |
Autonomous agent engine. A 4-layer pipeline routes ~80% of queries with 0 LLM calls and keeps measured per-session cost around ~$0.0005.
| Spec | Detail |
|---|---|
| Retrieval layer | ChromaDB + SentenceTransformer RAG, ~30ms latency (measured on 200-query benchmark set) |
| Cost control | Token tracking and daily quota enforcement |
| Pipeline | Planner to Validator to Executor to Replanner (on failure) to Responder. Auto-retry with replanning |
| Traceability | Every decision logged at session level for full audit trail |
| Stack | Python, LangChain, ChromaDB, Gemini, FastAPI, Pydantic |
| github.com/harshbhanushali26/ai-agent-engine | |
Full-stack AI finance platform. A dashboard for passive logging plus a conversational agent for active querying, both writing to the same SQLite store. Third generation of the system.
| Spec | Detail |
|---|---|
| Query routing | ~60% of queries answered by pattern matching, 0 LLM calls. Complex queries stream via Groq + SSE |
| Features | Financial health scoring, spending forecasts, budget threshold alerts |
| Auth | bcrypt-hardened auth with session tokens and CLI-based recovery |
| Stack | Python, FastAPI, SQLModel, Groq, SSE |
| github.com/harshbhanushali26/finos | |
6-agent supervisor-driven CLI pipeline. Search, scrape, summarize, critique, synthesize. Complexity-based model routing with Groq and Gemini fallback. Covers 30 to 50 sources across 5 to 9 targeted questions per run.
| Spec | Detail |
|---|---|
| Agents | Supervisor to Search to Scraper to Summarizer to Critic to Synthesizer |
| Model routing | Groq (llama-3.3-70b) primary, Gemini (gemini-2.0-flash) fallback |
| Search | DuckDuckGo search plus web scraping per query. Auto-detects research gaps mid-run and re-searches |
| Stack | Python, Groq API, Gemini API, DuckDuckGo API |
| github.com/harshbhanushali26/research-pipeline | |
Automated resume-to-JD gap analyzer. Full skill mapping and personalized roadmap in under 5 to 8 minutes. Clean parse, analyze, output pipeline.
| Spec | Detail |
|---|---|
| Parsing | Extracts unstructured resume data, evaluates against a raw JD or a job-post URL |
| Mapping | Correlates identified skill gaps to step-by-step learning modules |
| Pipeline | Parse to analyze to output. No hardcoded resume or JD assumptions, works on any format |
| Stack | Python, FastAPI, Groq API, Gemini, Tavily, Pydantic, REST API |
| github.com/harshbhanushali26/NextSteps | |
| Sheet | Function | Stack |
|---|---|---|
| DevMind, MCP Server | Local MCP server with 6 tools. File read/write, Python execution, JSON utilities. HITL confirmation on all write and run actions. Fixed a subprocess stdin isolation bug that silently hangs most naive MCP implementations. | Python, MCP SDK, tiktoken |
| LangGraph Parallel SQL Runner | Multi-question parallel SQL execution via LangGraph's Send API, with an inline human review gate before any query fires. | LangGraph, Groq, SQLite |
Tools and frameworks are grouped by the engineering capability they support — orchestration, retrieval, backend systems, reliability, and interfaces.
AI engineer with 9 months of full-stack production experience and a portfolio of agentic, RAG, and tool-execution systems. Open to AI Engineer and Python Developer roles, especially at teams building real AI products.
| LOOKING FOR | AI Engineer, Python Developer |
| AVAILABILITY | Open to AI Engineer roles at AI-first startups — Mumbai, Pune, Bangalore, Hyderabad, or remote. |