AI Engineer · San Diego, California
Ketan Shukla
I build production LLM and agentic systems — and the guard rails that make them safe to hand a dangerous tool.
Most of my recent work comes back to one question: what does it actually take to let a model do things, safely? Not demonstrations — hosts that own the loop, gate the irreversible actions, and can prove afterwards what happened and what it cost.
That turns out to be less a machine-learning problem than a systems one. Trust boundaries, idempotency, durable execution, release engineering, and knowing what a run spent. I spent the first part of my career on enterprise integration and release management for federal and national-retail clients, and it is the same problem wearing different vocabulary — which is a thing I have written about at length.
Everything here is deployed and open source, with live endpoints, regression suites and measured tradeoffs. Where I quote a number, it came from a real run.
The arc
Four eras, most recent first. The through-line is the same in all of them: get work through a multi-step process across systems you do not control, and know what happened when it does not arrive.
Agentic AI
2026
Systems that can be trusted with a dangerous tool
Six deployed Model Context Protocol projects, built in sequence, each solving a limitation found in the one before: a server, a host that owns the agent loop, a human-in-the-loop approval gate, sub-agent delegation, and a spend ledger for model sampling. Five further production AI applications across retrieval-augmented generation, agentic research, and structured extraction. Every claim carries a measurement from a real run.
- 0 of 36 targets for 439k tokens alone · 36 of 36 for 248k delegated
- 13 sampling calls: 8 allowed at 13.0¢, 5 refused at $0.00
- one cache breakpoint, 47% off input cost
Metronagon Media
2024 — present
An AI generation pipeline that shipped 22 books
Founder and sole engineer of a five-application production stack: Stripe embedded checkout, an end-to-end generation pipeline across the Anthropic API, GPT Image and Azure Neural TTS, and an automated book pipeline producing EPUB, PDF, DOCX and M4B audiobook outputs with per-run token cost tracking. Twenty-two books across three series, 388 chapters, published and distributed.
- 22 books · 3 series · 388 chapters
- 13 commercial fiction genres in the pipeline
- 5 live Next.js applications on custom domains
Data engineering
Python · data
Pipelines that fail loudly and validate at the boundary
ETL work across financial market data, COVID-19 integration, e-commerce sales, log analysis and a configuration-driven warehouse framework — separate extractor, transformer, validator and loader layers, transaction-safe loading with rollback, and mock servers for test isolation.
- 8 input source types in one orchestrator
- regex extraction at 97% parsing accuracy on unstructured logs
- RSI, MACD and Bollinger indicators verified against reference data
Enterprise integration
Earlier career
Federal message orchestration, and nightly builds for 400 stores
Enterprise integration on BizTalk Server for Citizenship and Immigration Canada at Avanade — receive and send pipelines, XSD-validated maps, and orchestrations moving applicant data through a multi-stage federal process. GIS automation for Canada Post. Before that, Point-of-Sale software across Giant Tiger stores nationally, where I ran the nightly build cycle and the after-hours release distribution, with a rollback path per store — in an environment where a POS failure stopped checkout.
- suspended-message diagnosis across schemas, maps and orchestrations
- national release distribution with per-store rollback
- Oracle and Sybase client/server, tuned for retrieval
What I write about
15 posts, four subjects, every one out of a deployed project.
Model Context Protocol
Retrieval-Augmented Generation
Agent-to-Agent
Model APIs & Application Surfaces
Elsewhere
Other things I run, in case you arrived here looking for one of them — or for a different Ketan Shukla.
- ketanshukla.com
- My author catalogue — 22 self-published books across three series, with the pipeline that produced them.
- metronagon.com
- Metronagon Media: book publishing services, cover and series branding, production pipelines.
- surplusrecoveries.com
- Helps former homeowners claim surplus funds left after a foreclosure sale.
Open to full-time, contract, and advisory work in AI infrastructure, agent platforms, and developer tooling. Based in San Diego, working worldwide. developer@ketanshukla.dev