Work

Ten things I’ve built, newest first, each with what was true when I got there and what followed.

  1. 01GARDE1
  2. 02Genomica and Nutrimentica
  3. 03Litehouse
  4. 04OC Sports & Entertainment
  5. 05Allē, for AbbVie
  6. 06Pangea
  7. 07Hoopla
  8. 08LabRecruit
  9. 09Thought Catalog
  10. 10Elexa

01

GARDE1

Founder. 2025 to now.

garde1.com

The tools handed people the work. They needed the work done.

CMMC tooling was stuck in 2010: checklists, and software that needed an expert to use it. 110 controls, 320 assessment objectives, and defense manufacturers with no security staff.

  1. Started with the competition. Vanta’s interface is its data model: everything bolted on, nothing specialized or guided. Delve promised agentic evidence gathering the benchmarks said wasn’t ready. Conway’s law applies: a product mirrors the org that built it. Vanta, Delve, Conway’s law
  2. Bet against browser agents. In 2025 the best computer-use agent finished 38% of the desktop tasks people finish 72% of, slowly and at real cost. That’s on a benchmark, where scores get optimized; real customer environments are worse. Evidence an auditor will test can’t work that way, so GARDE1 collects it directly. OpenAI, 38.1%, OSWorld, 72.36% for people, tens of minutes per task, cost and benchmark overfitting
  3. Turned 27+ menu items into states. Ease of use is the product, so the interface follows where you are in the work, not what the software contains.
  4. Built all of it. Thesis, brand, design, marketing and engineering. 8,731 commits across a .NET agent, a Go backend and a TypeScript frontend.

A mock CMMC assessment that runs end to end, and three inventions inside it. Two are patent pending.

Sources: CMMC, 110 controls, 320 assessment objectives (NIST SP 800-171A)

The teardown

Before building, I took the category apart. My read of each, and what GARDE1 does instead.

CMMC tooling

Where it breaks

Stuck in 2010. Checklists, and software that assumes you already understand 110 controls and 320 assessment objectives.

GARDE1 instead

One guided line through the whole assessment. The customer never needs to know the framework.

Vanta

Where it breaks

The interface is the data model. Everything is a bolt-on; nothing is specialized or guided for the job in front of you.

GARDE1 instead

The interface follows the states of the work: where you are, what’s next, what’s blocking you.

Delve

Where it breaks

Sold agentic evidence gathering. The benchmarks said that wouldn’t work yet, and in our evaluation, before any of it was public, it came with an RPO, a human consultant, attached. In 2026 Delve was publicly accused of fabricating compliance evidence, which it denies, and parted ways with Y Combinator.

GARDE1 instead

Evidence collected directly, by connectors and an endpoint agent, because browser agents can’t yet be trusted with evidence an auditor will test.

Drata

Where it breaks

Redesigned its own experience to cut friction across the same sprawl. Conway’s law: a product mirrors the org that built it.

GARDE1 instead

Designed around the work from day one, not the org chart, so there’s no sprawl to redesign away.

evidence 1, citedevidence 2evidence 3, citedevidence 4ruled outruled outchosennot reachednot reached

PWTL (Prompt Workflow & Task Language)

The model chooses an answer instead of writing one, cites the evidence it relied on, and keeps what it ruled out. It can’t invent an outcome, and uploaded files can’t instruct it. Patent pending.

f1f2f3f4f5factsrulesstatements

The GARDE1 document system

Written section by section from an organization’s facts. Every sentence can answer an auditor’s “why is this here?”, nothing is invented, and the same facts make the same policy. Patent pending.

you stay herewhat you needconfirm firstThe chat points the way through the product. It never takes you there.

Locus

A protocol that gives AI a map of the website. Customers who run machine shops, not IT, get around the product by chatting, and the chat, reading the map, points the way without moving them. It sees only what a user would see: the page they’re on, and the other pages rendered quietly in the background. It stops at confirmation for anything sensitive.

Read the full story

Genomica

02

Genomica and Nutrimentica

Coming soon.

The guess isn’t the clinician’s fault. It’s the missing map.

Psychiatry’s first prescription is a guess. About one in four people with depression reach remission on their first antidepressant.

  1. Came to it as a biohacker. An engineer interested in biology, I knew the category leader well. SelfDecode was revolutionary for what it does, then it became the everything report: all noise, no signal. Everything was an add-on, and none of it was close to clinician-ready. SelfDecode
  2. Named the failure it prevents. Picture a psychiatrist prescribing Wellbutrin, which raises dopamine and norepinephrine, to a patient whose COMT gene clears those same signals slowly. The result can be irritability, agitation and insomnia. Genomica is built to flag that mismatch before the first prescription, not after. bupropion, COMT
  3. Started at the beginning of the pipeline. Doctors often act on averages, at the end of the pipeline: large trials wash out an effect like slow COMT, and drug studies on single genetic variants are rare. Genomica starts at the beginning, with the person and their genome, and gives doctors a plausible, auditable path from known mechanisms to a decision. averages hide individual effects, gene–drug guidance
  4. Treated pathways as graphs. From genome, to mechanism, to drug, with stochastic flux across the pathways in between. biological networks, metabolic flux
  5. Built the proof of concept. A real clinical decision system, with an auditable path physicians can bring into medication management. In Rust, powered by 20 GB of open-source genomic data.
  6. Added the other half of the medicine cabinet. Nutrimentica, a DrugBank for supplements, ties what each supplement does in the body to its effects.

Genetic clinical decision support for psychiatrists, with supplements in the same map as drugs.

Sources: one in four reach remission first time (STAR*D reanalysis, BMJ Open 2023)

The teardown

The tools a patient or a psychiatrist can reach for today, and what Genomica and Nutrimentica do instead.

SelfDecode

Where it breaks

Revolutionary for what it does, then it became the everything report: all noise, no signal. Everything is an add-on, and none of it is clinician-ready.

Instead

Genomica: one clinician-ready path, from genome to mechanism to drug, with the reasoning auditable.

GeneSight

Where it breaks

The best-known psychiatric genetic test. It mostly reports how a patient metabolizes drugs.

Instead

Genomica: models the pathways behind the symptoms, not only metabolism.

Examine

Where it breaks

Summarizes the research on each supplement, outcome by outcome. It doesn’t tie that research to mechanism, or to a person’s genome.

Instead

Nutrimentica: ties what each supplement does in the body to its effects.

DrugBank

Where it breaks

The reference for drugs and their mechanisms. Supplements, which many patients take alongside their prescriptions, aren’t its focus.

Instead

Nutrimentica: a DrugBank for supplements, in the same map as the drugs.

Read more

03

Litehouse

Staff product engineer. 2024 to now.

People needed answers, not query languages, and permissions that held.

A smart-district platform, proven at the Honda Center, becoming a commercial product.

  1. Rebuilt the natural-language BI. PwC’s $6M delivery shoved the whole GraphQL schema into the prompt. A temporal engine, a trimmed schema, a skills-based semantic layer and an agentic loop replaced it, and took it from 70% to 96% over a curated set of questions. the full story
  2. Made one platform serve many. Designed the multi-tenant system, multi-application too, and redesigned 40+ domain relationships.
  3. Federated the data. Architected the federated GraphQL service across the data warehouse and the application data.
  4. Invented the framework. An in-house framework on top of NestJS that adds MSSQL tables to the platform mechanically, instead of by hand.
  5. Made it safe to sell. SOC 2, Auth0 instead of home-rolled login, secrets centralized, and an access-control proxy in front of the AI tools.
  6. Shaped the product with the VP of Product. Refined what the platform did and how it looked, together.

Answers that respect who’s allowed to see what.

04

OC Sports & Entertainment

Senior product engineer, for the Anaheim Ducks’ owners. 2022 to 2024.

ocvibe.com

Prove an architecture can exist before you bet a launch on it.

A smart district being built from scratch, on Aegis: open-source module federation for the backend, so new it was still being proven.

  1. Tested the bet. Audited whether backend module federation could work, then wrote the roadmap for Aegis’s stability and developer experience.
  2. Delivered anyway. When it was clear Aegis wouldn’t be ready for the Ducks’ new fan-experience app, I pivoted to a stand-in, trained 30+ engineers across the US and India on it, and we shipped on time.
  3. Set the bar, and held it. Authored the house code guide, and reviewed every pull request, 1,000+ of them, refusing to merge until it met the bar.
  4. Grew the team. Mentored an offshore team of 30 engineers through technical 1:1s.
  5. Designed the backend platform. The architecture and the framework, the multi-tenant schema, authentication and authorization, and the pipeline that pushes live NHL game notifications, with the live-scoring integrity problems worked out.
  6. Led a team of three. A KPI dashboard for the Chief Innovation Officer.

The Ducks’ fan experience, on time, and the backend the district’s products are still built on. Aegis went on to become the foundation of TAHO, an execution fabric for GPU compute.

Sources: Aegis on GitHub, TAHO

05

Allē, for AbbVie

Senior software engineer. 2020 to 2022.

alle.com

2M+monthly visitors searching with it

Search and schema discipline were the product, not features of it.

A new booking and rewards platform. No real search, and a federated graph of 20+ services with no shared rules.

  1. Knew when Postgres wasn’t enough. Leadership wanted Postgres. Postgres is great, and the old adage holds: use Postgres until you can’t. Provider taxonomies were bigger than it could search comfortably, so I convinced leadership to bring Elasticsearch to AbbVie, then built the search service behind alle.com. The adage
  2. Built the core. The Provider service in Go, the one most of the platform depended on.
  3. Gave the graph a government. Helped voice the need for a GraphQL Council, then co-founded it; it reviews every schema change.
  4. Made the case for Go. Taught it to 60+ engineers, because a large team ships faster when the code is predictable.

The primary search on alle.com, for over two million visitors a month.

06

Pangea

Advisor to the CTO. Y Combinator W21, 2021 to 2022.

pangea.app

99.9%uptime, from several crashes a day

The crashes were the symptom. The missing practice was the cause.

A fast-growing marketplace with a non-technical founding CTO: memory leaks, crashes and slowdowns from a WebSocket used as an API, an unsupported database, and calls that should have waited.

  1. Coached the CTO. Architecture, infrastructure and practice, directly.
  2. Stopped the crashes. WebSockets to HTTP, slow calls onto queues, RethinkDB to Postgres.
  3. Taught the practice. Made the case for tests and pull-request reviews, and helped the team make them how work got done.
  4. Raised the floor. TypeScript, linting and tracing, so fewer bugs shipped.

From several crashes a day to 99.9% uptime, and a CTO who could run it alone.

Read more

07

Hoopla

Senior software engineer. 2020.

hoopla.net

What a team celebrates is what it does more of.

Software that helps sales teams recognize their wins.

  1. Let colleagues cheer. A new “fan” role, carried across 13 microservices.
  2. Met people where they work. A new service that delivers recognition into Microsoft Teams.

Wins, birthdays and anniversaries, celebrated inside Teams.

08

LabRecruit

Consultant. London, 2019.

Built a pipeline good enough for them to sell.

A recruiting firm hunting rare functional programmers, and job boards came up empty. The best engineers’ résumés were their code.

  1. Searched the code, not the résumés. Found engineers through their public commits in the target languages.
  2. Turned emails into people. The step the whole pipeline hinged on: reverse-engineered LinkedIn Sales Navigator’s API to turn commit emails into names and profiles.
  3. Made it searchable. Verified contacts, matched profiles, one index.

Recruiters searching by what people had actually built. LabRecruit went on to sell the pipeline to another agency. It’s the same idea today’s AI recruiting tools run on.

Read more

09

Thought Catalog

Software engineer. 2017 to 2020.

thoughtcatalog.com

300%revenue growth in my first year, on a platform I kept fast and cheap

Growth only pays if the platform stays fast and cheap as it grows.

A media company with millions of monthly readers, growing fast.

  1. Turned goals into products. With the Director of Innovation.
  2. Made it cheaper. AWS costs down 40%, with no downtime.
  3. Made it faster. An 85% faster database and 60% faster pages.
  4. Built the money paths. Recurring payments and a recommendation engine.
  5. Rode the trends with the archive. Paired Elasticsearch with Ahrefs search data to find which existing content matched a trend, like how to segment a library of quotes, so the team could hop on it fast.
  6. Put the newsroom in Slack. A Slack bot that ran the editors’ workflows in Slack, replacing work staff had done by hand.

The company’s revenue rose 300% in my first year, and the platform kept pace. 300+ deploys.

10

Elexa

Full-stack developer, contract. 2017.

At the moment of purchase, the checkout is the product.

Guardian leak prevention: a home-automation hardware company that needed to sell online.

  1. Made checkout one page, one step. Designed and built a frictionless checkout on a headless CMS, with nothing between the customer and done.

“The easiest checkout ever,” in customers’ words.

Smaller things

Made because they needed to exist.

yake-rust

A port of YAKE!, the statistical keyword extractor from researchers in Portugal, to Rust and WebAssembly, so Node.js could have one that runs at native speed.

crates.io, npm, the YAKE! paper (Campos et al., 2020)

Universal iCal for Tidbyt

A calendar app for the Tidbyt retro display that works with any calendar, written in Starlark.

Merged pull request, App, Backend

Seven Tests for Consciousness

Humans defined consciousness, so any test for it has to start from the human experience.

Read the paper