I treat trust as an engineering problem.

Fractional CTO and product engineer, Las Vegas / New York. I build products solo end-to-end, or architect them and lead the team that ships them. Full case studies at koltadams.com — this page is the engineer's cut.

As of Aug 2026.

Seasonal cooking platform. Solo, six months, 1,500+ commits. CI/CD on GitHub Actions, zero-downtime deploys. Source is private; the product is not.

• The site rotates itself. A seasonal config drives homepage, navigation, produce guides, and editorial by month. No manual updates.
• 20-job intelligence network, not 20 crons. Jobs trigger and inform each other through a dependency graph with tracked causal links (SEO audit → description polish; trending detection → auto-feature). Content lifecycle uses 7d/30d engagement moving averages and linear-regression decay to predict unfeature dates.
• Every AI write is gated. All proposals land in one review queue with side-by-side diff, Levenshtein-based change-magnitude confidence scoring, and a full audit trail. Apply writes to the database only after human approval.
• Two-stage recipe entry. Sonnet parses raw text in ~2 s; Opus classifies across 100+ filters under strict only-what's-written rules.
• Search. Ingredient-aware with synonym groups, plus Claude-interpreted natural-language queries compiled to structured filters. Sub-50 ms via PostgreSQL GIN indexes.
• Platform. 180 KB production bundle (NYT Cooking ships 380). 60+ normalized tables. Docker + GitHub Actions zero-downtime deploys. Redis, Elasticsearch, New Relic.

React 19 · TypeScript · Node · PostgreSQL · Redis · Docker · Claude · Gemini

Native iOS supplement and longevity tracker. Private beta.

• Content is AI-drafted, clinician-gated, schema-enforced. A database-level constraint makes it structurally impossible to publish a high-severity interaction warning backed only by preliminary evidence. The invariant lives in the migration, not in reviewer discretion.
• 100 published interaction records across 41 substances, each reviewed by a licensed physician. Seven of 107 were retracted and the systematic errors behind them fixed at the source. Pharmacist completeness review gates public launch.
• Nightly biometric ingestion from Oura and Apple HealthKit, with per-source retention honoring each provider's data terms.
• Evidence-graded — every claim carries its evidence level, and the UI shows it.

SwiftUI · HealthKit · Node · PostgreSQL · Cloudflare Access · Claude

Mental-health screening orientation. Helps someone decide whether what they're feeling is worth a professional conversation. Never a diagnosis.

• Two model stages. Free-text routing to a validated screener (GAD-7, PHQ-9, PCL-5, others) or straight to crisis support; then plain-language, non-diagnostic reflection of results.
• Eval-driven. 46-profile classification suite at 100%, with a crisis-detection floor that held through every prompt revision and 42 out-of-scope guards against misrouting. Prompts that didn't clear the floor didn't ship.
• Accountless, on-device, zero PII. Crisis support one tap from every screen.

React Native · Expo · TypeScript · Cloudflare Workers · Claude

Client work under a hard identity-guideline brief. No framework, no analytics, no cookie banner. Static files, a Makefile, and a Playwright screenshot-diff verifier so a stylesheet change can't regress a page nobody looked at.

Zero-dependency CLI. Scans source for process.env usage, diffs against .env.example, reports missing, undocumented, and unused vars. Exits 1 on issues, so it drops into CI.

[email protected] · koltadams.com · LinkedIn