Fractional CTO & product engineer

I treat trust as an engineering problem.

Architecture, design, and code — for founders who can't afford the wrong hire. Evidence over adjectives; safety encoded in the schema, not left to good intentions.

SystemStateEvidenceAs of Aug 2026

koltstable.com live 180 KB bundle · 96 Lighthouse · search <50 ms toren private beta 100 interactions · 41 substances · clinician-gated inkling TestFlight 46/46 eval profiles · 0 personal data lopudmanagement.com live static · 0 runtime dependencies · client work

Four systems, four decisions.

Descriptions are cheap. What a CTO actually sells is judgment, so each entry below is the one hard call I made, what it cost, and how you can check it.

Kolt's Table

Seasonal cooking platform. Solo, six months, brand through infrastructure.

Problem: Seasonal food content goes stale in weeks. One person can't keep a site current by hand, and a site that isn't current is worthless.
The call: Make the site rotate itself. A seasonal engine turns the homepage, produce guides, navigation, and editorial over by month. Behind it, a 20-job intelligence network — jobs trigger each other through tracked causal links, not independent crons — drafts, scores, and proposes. Every AI write waits in a human approval queue. Nothing touches the database without me.
What it cost: A 60-table normalized schema and a governance layer before the first recipe published. Slower to start; nothing to unwind later.
Proof: Production bundle 180 KB where NYT Cooking ships 380. Search responds under 50 ms. Lighthouse 96. Natural-language recipe search and ingredient matching run on Claude in production today.

Toren

Native iOS supplement and longevity tracker. Private beta.

Problem: Supplement-interaction data online is confidently wrong. A tracker that repeats it is worse than no tracker.
The call: 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 safety invariant lives in the schema, not in reviewer discretion.
What it cost: Every record read by a licensed physician; 7 of 107 retracted, and the systematic errors that produced them fixed at the source. A pharmacist completeness review is a hard gate before public launch, not a nice-to-have.
Proof: 100 published interaction records across 41 substances, each human-reviewed. Nightly biometric ingestion from Oura and Apple HealthKit, with per-source retention honoring each provider's terms. In private TestFlight with an active household of testers.

Inkling

Mental-health screening orientation. Helps someone decide whether what they're feeling is worth a professional conversation.

Problem: A model that guesses wrong here does harm in both directions — missing a crisis, or medicalizing a bad week.
The call: Two model stages, neither of which diagnoses. The first reads a person's own words and routes them to a validated screener — or straight to crisis support when the language calls for it. The second reflects results in plain language that points toward care. Accountless, on-device, no personal data collected.
What it cost: A 46-profile evaluation suite with a crisis-detection floor that had to hold through every prompt revision, plus 42 out-of-scope guards against misrouting. Prompts that didn't clear the floor didn't ship.
Proof: 46 of 46 profiles passing. Crisis support one tap from every screen. Shipped to TestFlight.

Lopud Management

Client work. A talent firm with one client, whose studio dictates exactly how her name may appear.

Problem: The brief was restraint: expensive, weightless, nothing corporate — under an identity-guideline document from her team that vetoes shadows, recolors, and overlap.
The call: No framework. No analytics, no cookie banner, no about page. Static files, a Makefile, and a screenshot-diff verifier so a stylesheet change can't regress a page nobody looked at. Her signature machine-traced from the original scan, cleared by her team.
What it cost: Deleting a Vite/Tailwind scaffold that was never doing anything, and rebuilding deploy around what the host actually supports.
Proof: Live, indexed, and pixel-verified against its own previous release on every deploy.

Two ways to hire me.

Fixed scope, fixed price, written down before we start. No discovery phase, no hourly meter.

V1 build

You have an idea, a deck, or a codebase you don't trust. In ten to twelve weeks you have a shipped product, a documented architecture, and infrastructure in your name. Design and engineering, one person.

Handoff is the deliverable. You own everything. Thirty days of post-launch fixes included; after that, a retainer if you want one.

From $45,000 · fixed, agreed before the first commit

Fractional CTO

For teams that already ship. Architecture, the hard calls, and the hires. Week one you get a written assessment of what's load-bearing, what's rotting, and what to do first. No deck.

Two days a week, month to month, three-month minimum.

From $12,000 a month

How it starts

A thirty-minute scoping call. You tell me what has to exist at launch and when. Within twenty-four hours you get a one-page proposal: scope, timeline, price, start date. You say yes or no.

Every engagement I've led has shipped on time.

Taking on new work · Next start September

Book a scoping call ↗

or [email protected]

LinkedIn GitHub Las Vegas · New York