AI Systems Engineer
I build AI systems that reason, use tools, and operate in real business workflows.
Agentic AI · LLM Systems · Automation
With humans kept in control.
Open to AI Engineering, Agentic AI, AI Solutions & Forward Deployed Engineer roles
System status
Online- Portfolio
- Online
- GitHub
- Shubham261299
- Projects
- 7
- AI experiments
- Active
- Last build
- Exploring
- Agent evaluation, MCP security
“Turning LLMs, tools and data into real-world AI systems.”
- Years building production automation & AI
- 2.8+
- Automation frameworks shipped and run
- 10+
- Transactions / month handled
- 3,000+
- Manual effort reduced
- ~80%
- Invoice extraction accuracy (10 → 2 min each)
- ~95%
- Exception-handling effort cut with LLM decisioning
- ~40%
Featured systems
Real systems. Real impact.
AI agents, automation and intelligent systems built for actual business workflows. Pick a system to see the problem, the architecture and what was engineered.
01 / 07|Multi-agent · LangGraph · Human-in-the-loop
Agentic Case-Routing Orchestrator
A reusable orchestration engine that routes support cases to pluggable AI agents, with a critic, bounded re-planning and a human approval gate.
The problem
Support cases need different investigations, and teams want to plug in their own agents without rebuilding orchestration, state or approval logic each time.
The result
- Domain agents plug in behind one contract
- Concurrent case processing
- Durable checkpoints: approvals survive a restart
- Bounded re-planning keeps runs predictable
System architecture
- Case inputEmail · Form · API
- Router agentClassify & route
- Agent AAgent BAgent C
- Critic agentValidate & review
- Human approvalReview & approve
- Final outcomeCase closed
Execution (simulated)
- Analyzing casedone
- Routing to a domain agentdone
- Agent processingdone
- Critic validatingdone
- Awaiting human approvaldone
- Case closeddone
State (simplified, illustrative)
{"case_id": "C12345","state": "routing","agent": "domain_agent_a","confidence": 0.92,"human_approval": "pending"▍}Experience
Automation → Intelligent Automation → GenAI → Agentic AI
Each step added a layer: reliable automation first, then documents and APIs, then LLMs, then agents that act with human oversight.
Career evolution
From data and automation to building agentic AI systems.
- 2019Data SciencePython, SQL
- 2023AutomationRPA, Power BI
- 2024Intelligent AutomationOCR, APIs
- 2025GenAIRAG, LLMs
- 2026Agentic AILangGraph, MCP
- NextAI Systems · FDECustomer-facing AI
Now
Agentic systems with human oversight: LangGraph, MCP and local LLMs.
Next →
Customer-facing AI work: Forward Deployed Engineer and AI Solutions roles.
Engineering philosophy
Principles I follow while building agentic AI systems.
- 01Ground the modelFacts before generation.
- 02Give agents toolsIntelligence becomes useful when it can interact.
- 03Keep humans in controlAutomation ≠ removing accountability.
- 04Make state durableAgents should survive failures.
- 05Measure the systemEvaluation > vibes.
- Jul 2026 – NowNow
Agentic AI
Automation Engineer · IDeaS Revenue Solutions, Pune
Building agentic AI and automation inside a business team: multi-agent orchestration with human approval, MCP servers for controlled agent access to enterprise systems, and local-LLM pipelines.
- Built a LangGraph multi-agent engine: pluggable agents behind one typed contract, a critic with bounded re-planning and a human approval gate. Paused cases survive a kill-and-restart
- Built two MCP servers (Salesforce read-only, Smartsheet controlled read/write) with fail-closed API-key auth, per-key rate limits, allow-lists and audit logging, deployed internally
- Hardened every LLM step: schema-constrained output with a repair retry and a deterministic fallback, plus timeouts so one failed agent never crashes a case
- Building a local-LLM report pipeline where code computes the numbers and the model writes only the narrative
- LangGraph
- MCP
- FastAPI
- Local LLMs
- Human-in-the-loop
- Dec 2023 – Jul 2026
Automation → Intelligent Automation → GenAI
Analyst · Searce, Pune
Owned automation end to end for logistics and accounts-payable workflows, from requirements to L2/L3 support, then added LLMs to it.
- Built and ran 10+ automation frameworks (UiPath, Selenium, Playwright, REST APIs) handling 3,000+ transactions a month
- Invoice document processing with OCR: ~10 → ~2 minutes each at ~95% accuracy
- Added LLM decisioning to exception handling, cutting that effort ~40%
- Built a local RAG assistant over SOP documents with page-level citations
- UiPath
- Playwright
- OCR / IDP
- REST APIs
- RAG
Education | 2019 – 2023
B.Tech, Information Technology (Data Science)
Ajeenkya DY Patil University, Pune
CGPA 7.85 / 10 · Data, Python and analytics fundamentals.
Early projects
Certifications
- Full Stack Generative & Agentic AI with Python — Udemy
- UiPath Automation Developer Professional Training — UiPath Academy
- Python — Kaggle
- Forage virtual job simulations: BI Developer (Apr 2023), Data Analyst (Dec 2022)
AI Lab
What I'm exploring and building
Experiments
Ideas, experiments and things I'm exploring.
- MCP SecurityAuthorization patterns for AI tools
- Agent MemoryPersistent agent state experiments
- Local LLMsRunning private AI infrastructure
- Agent EvaluationLLM-as-judge experiments
- Self-Healing AutomationMaking browser automation resilient
Currently building
In progressOpspilot
AI back office for small businesses.
- Excel
- AI processing
- Agent proposals
- Owner approval
Recent activity
View GitHubLive from GitHub (public repositories).
- Add Shubham AI Lab banner and links header to READMEdecision-intelligence-lab ·
- Add the policy search eval and document week 6opspilot ·
- Record Day 4 results and add a resume section to the planagent-blackbox ·
- Document week 5: tracing, evals and the MIT licenseopspilot ·
- Prepare DecisionLab software-first releasedecision-intelligence-lab ·
Stack
The stack behind the systems
- Agentic AI & LLMs
- LangGraph
- LangChain
- Multi-agent orchestration
- ReAct & tool calling
- Model Context Protocol (MCP)
- Human-in-the-loop
- LLM-as-judge evaluation
- Guardrails
- Models & Retrieval
- Claude
- OpenAI
- Gemini
- Hugging Face
- Ollama (local LLMs)
- RAG
- FAISS
- ChromaDB
- Embeddings
- Backend & Integration
- Python
- FastAPI
- Pydantic
- REST APIs
- SQL Server
- PostgreSQL
- SQLite
- Docker
- Automation & Document AI
- Playwright
- Selenium
- UiPath
- OCR / IDP
- OpenCV
- PyMuPDF
- Power Automate
- Data & Tooling
- pandas
- Power BI
- Git
- Linux
- Streamlit
About
Behind the lab
Based in Pune. I came up through data and automation, and what I care about most is AI that holds up inside real workflows: observable, recoverable, and with a person in the loop where it matters.
Most of what's on this page is built and run, not just demoed. The rest is in the lab, in public.
Let's talk
Let's build something impactful.
I'm open to AI Engineering, Agentic AI, AI Solutions and Forward Deployed Engineer roles.
shubhamjamdar2612@gmail.com