Grounded or refused
Generated claims trace to evidence or don’t ship. Gaps get acknowledged, not papered over. Enforced in the system, server-side, not in the prompt.
I build AI systems that can prove what they did, and refuse when they can’t.
Grounded generation, deterministic cores, eval-driven development, and a receipt for every action. Built end to end, run in production, operated solo.
19 years shipping software Orange County, CA Python · TypeScript · C++ LLM · RL · CV, in production
Interrogation desk, the thesis, live
Ask the record anything.
This bot answers only from a curated evidence corpus and must cite it. The server verifies every citation, an answer it can’t back gets stamped refused, not improvised. That constraint is the same architecture the systems below run on.
Grounded-or-refuse · citations verified server-side№ live
Ask about the systems, the incidents, the method, or the 19 years. If it isn’t in the record, I’ll say so.
source: E53, this bot is itself the demonstration
Ledger, systems built end to end
The work.
Every card links to a full account: problem, architecture, the hard decision, the numbers, and what broke. Code is proprietary; the systems run, and I walk through any of them live.
Entry 001 · live in production
A job-search platform that never fabricates
RAG-grounded résumé tailoring with server-side enforcement, immutable receipts for every application, a self-calibrating evaluation funnel, graduated agent autonomy.
13 lifecycle states 27 agent tools 24 security findings fixed ~90s résumé→evaluation
Entry 002 · sensitive-data domain
Analytics that are never confidently wrong
Plain-English questions over regulated HMIS data. The LLM proposes typed intent; a deterministic engine computes; a validator gates every number, or the system asks instead of answering.
112k LOC 125 test files offline eval harness 0 ungated numbers
Entry 003 · systems + RL
Backtest and live are the same code
C++23 market terminal and strategy engine: mmap’d POD stores with sanitizer-verified pointer stability, offline RL with pre-registered experiments and champion–challenger gates.
147k LOC C++23 295 test files 39 ADRs ~150 RL checkpoints
Entry 004 · applied CV + optimization
A grocery agent with a fail-closed guardrail
Constrained basket optimization across cost, nutrition, and trips, LLM only for substitutions, with on-device computer vision, live retail-API integration, and a verified-savings ledger.
YOLOE/OWLv2 on-device CV fail-closed allergen gate live retail API
Record, 2007 to present
Nineteen years, in brief.
2025, present
Production AI systems, end to end
Designed, built, and operate the four systems above, solo. Product, architecture, implementation, deploys, backups, incident response, postmortems. LLM, RL, and CV workloads with evals, cost controls, and guardrails as first-class architecture.
2007, 2024
Client web development, every era of the stack
Paid web work from 2007 onward, through the PHP/jQuery years, the SPA turn, and into the modern TypeScript stack. The full account of clients and engagements is available on request; this public record keeps client work private by default.
2007
First paid build
Where the ledger opens.
Method, what holds across all of it
Four rules, no exceptions.
These aren’t aspirations; they’re load-bearing architecture in every system above, and in the bot at the top of this page.
Deterministic core, LLM at the edge
The model proposes; typed, testable machinery disposes. Numbers come from engines and validators, never from a language model’s mouth.
Eval-driven, or it isn’t engineering
Non-deterministic systems get golden sets, agreement measurement, and calibration loops that only ship a change when it provably helps.
Every action leaves a receipt
Immutable records of what was done, with what inputs, at what cost. Auditability is a feature users see, and the reason autonomy can be granted at all.
Writing, the reasoning, in public
Essays
Colophon
About
I’ve been paid to build software since 2007, client web work through every era of the stack, and for the last two years, production AI systems built end to end: retrieval pipelines, agentic workflows with guardrails, evaluation harnesses, cost controls, and the unglamorous operations underneath.
I work the full span: product decisions, architecture, implementation, and running the result in production with real users. The common thread is a distrust of unverifiable output, mine included, and systems designed so trust doesn’t have to be assumed.
I’m based in Orange County, California, and open to AI engineering leadership jobs.