Senior AI engineer in Pune. I build agents that do real work, and write down how they break.

Three years shipping production GenAI for manufacturing, automotive, sustainability and real estate clients: multi-agent systems, RAG, NL2SQL and fine-tuning on Azure AI Foundry and Microsoft Fabric.

Lately, from the commit log. It's in Pune.
  1. Sable release: 1.0.0 (#109)
  2. Sable test: Phase 9 gate run, all seven lines pass; v1.0 complete (#108)
  3. Sable feat(daemon): weekly self-check from existing data, no model calls (Phase 9 Task 7, K10) (#107)
  4. Sable feat(share): approve-only and read-only sharing through ntfy (Phase 9 Task 8, K12) (#105)
  5. Sable feat(hosts): run commands on other servers through their own Sable (Phase 9 Task 5, H5) (#104)
  6. Sable feat(memory): incidents become runbooks (Phase 9 Task 6, K9) (#103)

Client work

2025

Twelve agents that write workflows

One of India’s largest dealer management networks

Workflow creation went from 4 to 8 hours down to 2 minutes, and error rates from 30 to 40% down to under 10%.

Six workflow agents and five UI agents, from schema analysis to payload assembly, generate production workflow JSON of 500 to 2,000 lines and UI form configs from plain English, across 118+ task types, with handoffs and validated structured output.

Multi-agent orchestration, structured outputs, Azure AI Foundry

2025

Defect analytics that refuses plausible wrong answers

A global electronics manufacturer

Found a hidden join fan-out that inflated every test event 3.5x, rebuilt the grain, and compressed 1.53M rows to 145K with no loss of detail.

A medallion warehouse in Postgres with a governed semantic layer for NL2SQL. Known traps like cross-grain aggregation are hard constraints on the agent. The dashboard is a zero-backend React app querying Parquet in the browser via WebAssembly, deployable air-gapped.

Postgres, DuckDB, Parquet, React 19, WebAssembly, Docker

2025

NL2SQL across a six-tier dealer hierarchy

The same dealer management network

Natural-language analytics over 65,000+ job cards per dealer per month, with row-level security at every tier.

Medallion architecture on OneLake joining 23+ source tables from MySQL, SAP and REST APIs, queried through a Fabric Data Agent.

Microsoft Fabric, OneLake, Fabric Data Agent, Azure AI Foundry

  • 2024 to 2025Enterprise RAG over technical documentation for a global industrial valve maker, with custom table extraction and full observability.
  • 2024Agents that read Excel financial models and generate investor-ready PowerPoint reports for a real estate firm.
  • 2023 to 2024Llama 3 fine-tuned on cement and sustainability data, plus GraphRAG over Neo4j, in production for process optimisation.
  • 2023 to 2024Multilingual converter for work instructions into Indian languages, used by 7,000+ people across a manufacturing site.

Side projects

  • jevbriefClean, traceable state briefings for TypeSafe's Jev model.★ 4
  • ArchonEncrypted, verified database backups as a drop-in Docker sidecar.★ 3
  • VeritasChainHash a file, pin it to IPFS, anchor the fingerprint on-chain.★ 2

Writing

All writing

Experience

  1. 2025 to now
    Senior Software Engineer, Equations Work, Pune
  2. 2023 to 2025
    Software Engineer, Equations Work, Pune
  3. 2023
    Internship Trainee, Maxima Consulting
  4. 2019 to 2023
    BE in Computer Engineering, Genba Sopanrao Moze College of Engineering. Led competitive programming at the CodeChef chapter.

Certified in Model Context Protocol, introduction and advanced topics, by Anthropic (August 2025), and a Neo4j Certified Professional. Second of 30+ developers at the Metron Security Code Jam (January 2025), building an Elastic ML security dashboard in six hours.

Agents and LLMs
Multi-agent orchestration, RAG, GraphRAG, NL2SQL, MCP, Llama 3 fine-tuning, LangChain
Data
Postgres, pgvector, DuckDB, Neo4j, Parquet, medallion and semantic layers
Cloud
Azure AI Foundry, Microsoft Fabric, OneLake, Azure Functions, AWS, Docker
Code
Python, FastAPI, TypeScript, React 19, WebAssembly

Say hello

The fastest way to reach me is pkomalwad@gmail.com. I usually reply within a day.