Houston, Texas · forward-deployed AI lead · founder @ MidnightDev

I build AI products—
and the systems that get them discovered, measured, and used.

Before software, I worked in commercial real estate and contributed to 85+ shopping-center transactions totaling roughly $400 million.

Book a callSee how my agents workOpen the interactive showcase ↗hire me in-house →

I'm Alex Bouchard, a forward-deployed AI lead and AI-native founder-operator. At MidnightDev, I am the sole human accountable for every function and every production decision—from thesis, business model, brand, and UX through AI systems, full-stack engineering, testing, security, deployment, growth, analytics, and ongoing operations.

85+

shopping centers · Class C repositioning

≈2 weeks

concept to App Store approval

72 hours

ReadableByAI · concept to production

Agents

How I put AI agents to work

I give AI agents a goal and leave them to work. While I'm away, they can investigate, build, and test changes; another agent can review the result.

If the work is major, unverified, or outside its limits, it waits for me.

See how my agents work →read the essay →
  1. 01

    Goal

    I set what needs to happen.

  2. 02

    Work

    Agents investigate, build, and test the changes.

  3. 03

    Checks

    Automated tests, plus another agent that reviews the result.

  4. 04

    Release rules

    Only permitted changes can be released automatically. Major work waits for me.

Shipped products

Products that shipped—and the engineering evidence behind them.

The products are different by design. The case studies explain where the model helps, where it fails, and what the code keeps under control.

Interactive showcase

See two shipped products up close.

That's My Best and Yapword, from idea to the App Store, built by one person.

Open the interactive showcase ↗
Yapword daily word game board mid-round, with the persistent generative character reacting to the player's guesses.

Yapword

On the App Store

Live on web + the App Store, in English and Spanish · 2,300+ tests gate releases and character changes

Daily word game with a persistent generative character that reacts to real play.

reactgeminicapacitorweb + ios
That's My Best converting an uploaded photo-grid screenshot into a multiple-choice quiz about which friend picked which photo.

That's My Best

On the App Store

Live on web + iOS · concept to App Store approval in 14 days

Multimodal social game that turns photo-grid screenshots into a friend quiz.

multimodal aimachine-checkable invariantsweb + ios
Yapoleon's Court gameplay screen showing the AI opponent's in-voice reply beside a favor meter the model cannot write to.

Yapoleon's Court

Live on the web

Published judge calibration: 96.7 / 72.2 / 0 · 5 of 5 anti-gaming probes pass live

Competitive AI game with a calibrated judge that resists flattery, injection, and template farming.

reactgeminivercel
Dozen landing page with the headline Don't buy one influencer. Agents run a dozen, and a twelve-node campaign graph whose accept node is still running.

Dozen

On the web

113 production-imported tests (2026-09-01) · Stripe Checkout wired fail-closed · no GMV claimed

Agent marketplace where the model cannot accept, post, or pay.

tanstack startinvariant testsclosed-loop agents
+more work: Boardroom, Caduceus, Caliber
AI Boardroom Forecast Audit diagram showing six synthetic decision lenses reviewing a $65.9M revenue forecast and flagging it as unsupported.

AI Boardroom Forecast Audit

Reconstructed a $65.9M forecast down to a $1.22M median — publicly

Six synthetic decision lenses, one broken forecast, and a reproducible source-of-truth audit.

pythonmonte carloregressioninvariant tests
Casting matrix pairing each PrinciplesYou measurement with the council seat it produced — the Shaper least-like-Implementer archetype gap to FORGE, Organized 26 and Detail-Oriented 34 to FINISHER, Curious 95 crossed with Detail-Oriented 34 to RECON, Creative 94 and Feisty 91 to SPARRING, and Direct 94 crossed with Empathetic 41 to ENVOY — above the weighting formula, believability times confidence with a Laplace prior and 45-day half-life, and measured figures of 167 seconds and 14.7 cents per full council run with 62 passing tests and an outcome ledger not yet settled.

Caduceus

Single measured run (2026-08-19) · 5 seats, 4 model families · $0.147 total cost · 167s · zero seat errors

A five-seat AI decision council cast against a measured psychometric profile, with the strongest dissent published beside the recommendation.

pythonfour model familiesbelievability weightinghermes agent
Caliber concept diagram stamped concept, in design, 2026, showing three agent cards labeled doer, challenger, and finisher above calibration dials fed by a public-domain IPIP Big Five profile, with a note that the outcome ledger holds zero settled outcomes.

Caliber

Concept stage · no product, no users — and it says so up front

Design-stage concept for a three-agent team calibrated to a person's measured psychometric gaps rather than generic roles.

design conceptipip big fivethree-agent teamoutcome ledger

open the build room: the full case file →

What I build

Five kinds of work, one builder.

Business systems

Owned properties with a defined conversion path.

Useful tools, disciplined information architecture, organic acquisition, and a defined conversion path.

HTX Sell My Housenew · Oct 2026

↗

A Houston home-selling site for as-is and inherited houses. Its phone line is answered by Virtual Vinny, an AI voice agent that takes the owner's details, and a person calls back within one business day.

Citation Scoreboardnew · Oct 2026 · private use

The question after ReadableByAI: when buyers ask AI tools for a recommendation, is your business cited, and if not, why and what to change. A weekly measurement with one cause and one fix for every missed answer.

Texas Large-Load Trackernew · Oct 2026 · open source

↗

Watches ERCOT and Texas PUC filings on large-load grid hookups, mostly data centers, and writes a dated digest of what changed and which deadlines are coming.

BuyLandFL

↗

Florida land, statewide, county by county. Transactional search and programmatic county-level geography.

Search and AI visibility

I build the product. Then I engineer how it gets found.

A solo-built and operated portfolio across health research, real estate, legal information, local services, and consumer products. Every property below grows through organic search, with no paid acquisition.

ReadableByAIfree since Oct 2026

↗

Checks whether AI tools can read a website, taken from concept to production in 72 hours. Its first study, published on August 8th, found a quarter of Y Combinator's Fall 2025 startups invisible to AI, against 2% of established software companies. Weekly monitoring is now free.

BuyLandFL

↗

Florida land

statewide, county by county

Transactional search and programmatic county-level geography.

Methylene Blue Ultra

↗

Health research

comparisons, citations, calculators

Research-led topical authority, comparisons, citations, and calculators.

Jones Act Calculator

↗

Maritime injury law

tool-led legal discovery

Tool-led legal discovery with public methodology and intent-specific pages.

Midnight GEO Pro Pack: the commercial edition of the search-visibility skill suite, 16 operating skills for technical audits, topical mapping, indexing, analytics, and AI-answer readiness, licensed to a single seat.

+more owned sites: HTX Work Injury, HTX Permit Fix, HTX Foundation Fix, WordGameAI

HTX Work Injury

↗

A Houston industrial-injury information system combining guided qualification, calculation, and public safety data.

HTX Permit Fix

↗

A local-service acquisition system built around urgent permitting intent, guided intake, and recurring organic inbound calls.

HTX Foundation Fix

↗

A concept-to-live specialized business scaffold built in two days: acquisition, diagnostic, operating, financing, and local-discovery systems—not claimed contractor history.

WordGameAI

↗

An editorial comparison and methodology property that routes qualified discovery into the Yapword product ecosystem.

inspect AI citation research ↗inspect the distribution system →operated social surface: @YapoleonGreater ↗
Alex Bouchard, Forward-Deployed AI Lead and founder of MidnightDev, Houston, Texas.

About

I did not take the usual route into AI.

Before MidnightDev, I specialized in repositioning Class C shopping centers and contributed to 85+ shopping-center transactions totaling roughly $400 million. I worked across acquisitions, dispositions, leasing, and portfolio strategy—finding underperforming retail assets, diagnosing the broken economics, and creating value through disciplined execution.

That experience is why I treat model output as evidence, not authority—and why I model the business, acquisition path, and downside before calling a build complete.

Operating principle: Generate boldly. Validate cheaply. Kill ruthlessly. Scale what survives.

the build room: full case file →case file overview →resume →

Services

Start with the smallest engagement that can change the decision.

  • Product diagnostic

    A decision memo: what is true, what is assumed, and what to test first.

  • Proof-of-value build

    A bounded build that tests the hardest assumption, with a build, change, or kill gate.

  • Embedded engagement

    I work inside your team as a forward-deployed AI lead on a bounded problem.

Book a callall services →

Writing and ship log

What I learned, and what shipped this month.

all writing →ship log →

Contact

Have a product, growth, or operating problem worth attacking?

MidnightDev takes a limited number of consulting engagements, embedded product builds, and strategic partnerships. I am also open to unusually strong forward-deployed AI leadership, AI-native founder-operator, and special-projects roles.

alex@midnightdev.devHouston, Texas · willing to relocate
book a call →

or send a note — either reaches me directly

consulting services →

For technical reviewers

The engineering detail behind the work above.

Reliability and safety work, upstream contributions, research, measurement, and the order in which the experiments happened.

Reliability and safety

The model can generate content. It cannot own truth, scoring, safety, or spend.

Those boundaries appear in the apps and in small public reference systems for model fallback, content safety, and graph operations.

the canonical statement: the model cannot own the money →

Gemini reliability proxy

Fallback chains, retry budgets, output limits, and denial-of-wallet protection.

↗

LLM safety gate

Fail-closed classification with quorum voting and per-item degradation.

↗

Graphiti + Neo4j operations

Health-driven recovery, safe backups, and local-only networking.

↗

How I work with others

I work in other people's systems the same way I work in mine: evidence before opinion.

In getzep/graphiti (29k★), I production-validated the fix that removed a hardcoded OpenAI dependency: ran it on a live Neo4j + Gemini stack, caught it silently falling back to the provider it claimed to remove, and the error-handling change I suggested shipped in the merged PR. I keep an open RFC upstream on reranker configuration, a merged documentation PR in jashkenas/backbone explaining the Events internals, and co-author credit on a merged research PR in a 2.8k★ marketing-skills repo, where the maintainer folded my research into their own change — on the maintainer's terms, not mine.

Day to day I direct a multi-agent team. Disagreement is handled the same way I handle model output — as a claim that needs a receipt.

inspect the upstream record →

Research, search architecture, and measurement

Public research, reusable infrastructure, and commercial systems.

The portfolio extends beyond consumer AI: citation research, technical SEO tooling, independent invention work, and focused systems built around real commercial intent.

AI citation patterns

↗

A dated, source-qualified reference on how answer engines crawl, retrieve, and cite web content.

Midnight SEO skills

↗

An operational skill suite for audits, topical mapping, indexing, analytics, and search distribution.

Independent patent filings

Inventor on two confidential U.S. provisional patent applications filed in 2026 relating to controlled AI-system execution and human-gated authorization.

01

Search architecture

Technical SEO, content taxonomies, internal-link systems, programmatic geography, structured data, and indexation controls.

02

AI-answer readiness

Entity clarity, source-qualified claims, extractable answers, citation surfaces, freshness signals, and answer-engine research.

03

Owned distribution

Editorial properties, App Store positioning, social character accounts, conversion paths, and PostHog telemetry that closes the loop.

PostHog in production

Instrument the decision—not merely the pageview.

The measurement system connects product behavior, acquisition, reliability, and release decisions across the operated portfolio.

  • Five active PostHog project spaces across the portfolio
  • Dozens of explicit product events spanning web and iOS
  • Activation funnels, weekly retention, share-loop and referral measurement
  • Session replay, rage-click diagnosis, fallback health, and release telemetry

From experiments to product rules

Each experiment changed what I was willing to let a model control.

Dec 2025

Six AI perspectives challenged one forecast

Used deliberately different roles to stress-test a real event model. They found optimistic revenue assumptions; a later audit found a second error in how the source data was defined.

Jan–Feb 2026

Persistent agents created new failure modes

Gave agents memory, follow-up schedules, and delegated tasks. The result was useful, but made ownership and human review more important—not less.

Mar 2026

A 21-role simulation exposed bad handoffs

Simulated a small company to see where product, engineering, legal, and growth teams lost context. The failures pointed to shared state, written decisions, and explicit owners.

Mid-2026

The lessons became product rules

In Yapword, That’s My Best, and Yapoleon’s Court, models generate character and content; people and deterministic code control scores, confirmed answers, safety, and spend.

Aug 2026

Accept, post, and pay stayed human-owned

Dozen is a marketplace graph for 1k–50k creators. The model drafts captions. Tests own the gates. Stripe Checkout is wired fail-closed; Connect transfers and live posting are still scaffolding — the case does not claim GMV.