MidnightDev · Interactive case file · Services

Generate boldly. Validate cheaply. Kill ruthlessly. Scale what survives.

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

As founder and solo builder at MidnightDev, I work from product thesis and business model through UX, model behavior, full-stack engineering, web and iOS release, technical SEO, GEO/AI-answer readiness, owned distribution, and PostHog telemetry.

My specialty is the difficult seam where probabilistic models meet reliable software: persistent characters, multimodal generation, deterministic controls, invariant testing, safety boundaries, cost protection, analytics, and feedback loops. Models may propose and challenge; explicit authority boundaries and machine-checkable constraints decide what ships.

• Yapword · App Store
  AI-native daily word game where Yapoleon reacts to each guess, gives contextual hints, and comments on how the player actually played. Shipped across web and iOS with subscriptions, product analytics, and production regression monitoring.

• That’s My Best · App Store
  Multimodal social product that converts photo-grid screenshots into playable friend quizzes with generated questions, invariant-bounded answer keys, sharing, and group reveals. Taken from concept to App Store approval in 14 days.

• Yapoleon’s Court · Engineering case study
  Competitive character game where players try to earn Yapoleon’s favor. The model supplies bounded taste and personality; deterministic, server-owned code derives every score.

• PostHog across five project spaces — production event design, activation funnels, weekly retention, sharing and referral loops, paywall measurement, session replay, rage-click diagnosis, fallback health, and release telemetry.
• Owned organic-search portfolio — first-party Search Console snapshots include BuyLandFL at 51,804 impressions / 198 clicks over 92 days, Methylene Blue Ultra at 7,746 / 100 over 28 days, and Jones Act Calculator at 2,700 / 23 over 28 days—all organic, with no paid acquisition.
• Operated AI distribution — @YapoleonGreater on X, where the model drafts in character while owner-defined rules retain publishing authority and escalation control.
• HTX Foundation Fix — concept-to-live specialized business scaffold built in two days: acquisition site, diagnostic and risk-report concept, local SEO/GEO, intake, admin operations, financing/SMS architecture, and neighborhood targeting. A venture prototype—not claimed contractor operating history.

• That’s My Best: Model-Proposed Answer Keys Under Machine-Checkable Invariants — executable, sanitized reference model for answer-free player payloads, reveal only after a valid pick, immutable sealed state, and deterministic group reveal.
• Yapoleon’s Court — bounded character generation with prompt-injection isolation, safety filters, fallbacks, spend controls, and server-owned scoring.
• AI Boardroom Forecast Audit — reproducible case study showing how six synthetic decision lenses cut a roughly $65.9M forecast, followed by a source-of-truth audit that found a second $3.397M definition gap.
• gemini-reliability-proxy — model fallback, 429/5xx-aware routing, attempt and chain budgets, output-budget protection, and spend limits.
• llm-safety-gate — fail-closed, classifier-agnostic content-safety state machine with quorum voting, refusal/noise separation, and per-item degradation.
• graphiti-neo4j-ops — production-minded Graphiti/Neo4j operations with health-driven recovery, safe Community Edition backups, and local-only networking.

• AI Citation Patterns — dated, source-qualified field guide and reproducible benchmark for how AI search systems choose and cite sources.
• Midnight SEO Skills — versioned operating system that turns the citation research into auditable content, structured-data, and technical-SEO workflows.
• Independent patent development — completed a provisional patent application package for an independently developed invention, including the specification, prior-art review, examiner brief, and 26 claims, with critique from an Australian patent reviewer.

• Jones Act Calculator — interactive, search-oriented maritime compensation resource with public methodology and explicit legal disclaimers.
• HTX Work Injury — guided workplace-injury information system combining qualification logic, calculators, public safety data, and explicit legal disclaimers.

• getzep/graphiti #1637 (29.4k stars) — production validation on Neo4j 5.26 + Gemini; found the fallback path silently restored the OpenAI dependency the PR removed. The fix ships in factories.py.
• aaron-he-zhu/aaron-marketing-skills (2.5k stars) — co-authored the merged commit adding 2026 AI-citation guidance to geo-content-optimizer, upstreamed from AI Citation Patterns.
• getzep/graphiti #1698 — open PR for explicit MCP reranker-provider configuration.

Before software, I contributed to 85+ retail commercial-real-estate transactions totaling roughly $400M across acquisition, disposition, leasing, and repositioning. That background still shapes how I underwrite products, inspect assumptions, and make decisions under uncertainty.

TypeScript · React · Next.js · Capacitor · Python · Supabase · PostgreSQL · Neo4j · Docker · Stripe · Tailwind CSS · Vercel · PostHog

MidnightDev takes a limited number of product diagnostics, proof-of-value builds, embedded engagements, and strategic partnerships. I am also open to unusually strong forward-deployed AI engineering and applied-AI product roles.

midnightdev.dev · alex@midnightdev.dev · Houston, Texas · Willing to relocate