Agents that reason,
retrieve, and get things done.

I design and ship agentic AI systems: RAG pipelines, tool-using agents, and the evaluation loops that keep them honest.

View work Contact

Where the work has happened.

Jun 2026 - Present

Agentic AI Engineer

Deloitte USI

Building production agent systems: retrieval pipelines, evaluation loops, and the tooling that lets them run unsupervised.

Dec 2021 - May 2026

Systems Engineer

Tata Consultancy Services

Four-plus years in enterprise systems engineering, the production discipline that now shapes how I ship agents.

The build loop, every time.

Same five moves, whether it's a retrieval service or a warehouse copilot.

01

Frame the failure.

Before any orchestration code, map the ways the system can go wrong: bad retrieval, hallucinated citations, runaway tool calls, blown budgets.

02

Design the graph.

Model the workflow as a state machine, not a script. Nodes for retrieval, judgment, and correction. Edges for retries and human hand-offs.

03

Build the smallest agent.

Ship the thinnest version that could possibly work, then add tools, memory, and guardrails only as the evals demand them.

04

Grade every run.

An LLM-as-judge scores groundedness and relevance against a strict schema, with a human spot check on anything borderline.

05

Ship it observable.

Stream state over SSE, log every decision the graph makes, and leave behind a dashboard someone else can actually read.

Selected builds.

Five systems, five different failure modes to design around.

RAG-Ultra

A retrieval-as-a-service microservice with corrective RAG, contextual retrieval, and an LLM-as-judge fast path for high-confidence answers.

  • LangGraph
  • FastAPI
  • Corrective RAG
  • SSE
View repository
Document→Vision OCR→Chroma
Query→Condenser→Retrieve
Retrieve ⇄ Judge (retry on low confidence)
Judge: pass→Generate→Groundedness check
Answer + inline citations

SpendWise MCP

An MCP tool server paired with a live financial dashboard that watches the agent reason, call tools, and update budgets in real time.

  • FastMCP
  • LangGraph
  • SSE
  • Observability
View repository
SpendWise MCP agent observability dashboard, showing tool calls, budget health, and category breakdown

Agentic Task Manager

Turns screenshots, links, and voice notes into a prioritized todo list, with a human-in-the-loop gate before anything hits the database.

  • LangGraph
  • Gemini
  • FastAPI
  • React
View repository
Agentic Task Manager interface showing Kanban board and Eisenhower matrix

Project Assistant

A hybrid RAG assistant that keeps retrieval and embeddings local, and only calls the cloud for synthesis and grading.

  • LangGraph
  • Chroma
  • Ollama
  • DeepSeek
View repository
Retrieve→Generate→Evaluate
fail, loop < 3 → Web Search → Generate
pass or max loops→End
local Ollama embeddings→Chroma DB

SAP Warehouse Agent

A conversational cockpit that turns plain language into read-only SAP warehouse queries: stock levels, picking tasks, and inbound freight.

  • LangGraph
  • DeepSeek
  • SQLite
  • Streamlit
View repository
User→Auth→Agent
Agent→7 read-only SAP tools
stock, bins, inbound freight, outbound orders, picking
Async SQLite checkpoint→Response

What I actually spend my time on.

Agent orchestration

State machines built in LangGraph: conditional edges, retry loops, and human-in-the-loop interrupts for anything that shouldn't run unsupervised.

Retrieval architectures

Corrective RAG, contextual retrieval, reciprocal rank fusion, and layout-aware ingestion for PDFs, tables, and diagrams.

Evaluation and guardrails

Structured LLM-as-judge grading and groundedness checks, with a fast path for answers the system already trusts.

Tool use and protocols

MCP servers, function-calling agents, and tool integrations that talk to real systems, including SAP warehouse APIs and live budgets.

Serving and observability

FastAPI gateways, SSE streaming, and dashboards that show every decision the graph makes as it makes it.

Model flexibility

DeepSeek, Gemini, OpenAI-compatible endpoints, and local Ollama embeddings, swapped per task and per budget.

Working stack

Python LangChain FastAPI Gemini React TypeScript Docker SQLite Python LangChain FastAPI Gemini React TypeScript Docker SQLite

Have a hard agent problem?

Open to agentic AI engineering roles and contract builds.

Contact GitHub