Decisions, with receipts
Engagements and builds from the record. A link marked full story is my own account of what happened; a link marked receipt points at something you can go and check yourself. The recent builds are running in production right now, and the long versions live in the Field Manual.
Twenty-plus surfaces, one identity core, zero employees
The situation: 2026. A financial advocacy platform and its product family needed to exist fast: QR safety tools, e-signatures, independent research, fourteen vertical advocacy sites.
The call: One identity core on the Cloudflare edge, passkey-only, and build it solo instead of hiring first.
The result: 20+ live surfaces shipped from zero in months. Funded; the raise closed in weeks because the product was already running. The receipt is live →
The $46-a-month QR code, rebuilt free in 20 minutes
The situation: A friend was quoted $46 a month for a QR code that would stop working the moment the subscription lapsed.
The call: Build the first version in under 20 minutes, run it entirely in the browser so nothing leaves the device, and give it away.
The result: A few months later: 175+ QR and barcode types, a QR safety checker, a pro tier, and about 3 million requests a month, climbing. Hosting cost: zero dollars. Full story →
The vendor claimed 98.7% accuracy. The emergency room got 68%.
A speech-recognition vendor demoed 98.7% accuracy on scripted audio. Deployed in an emergency room, the real number was 68%. So we benchmarked on real radio traffic and real first responders instead of clean corpora and actors, and reported the stratified numbers even where they looked worse for us.
The published benchmarks came out lower than every competitor’s marketing page, and the real-world performance came out higher. That testing discipline is how I evaluate every vendor claim in a diligence. Full story →
Dozens of blockchain pitches. One question killed 95% of them.
The situation: 2017 to 2018, technical due diligence on dozens of blockchain startups at the peak of the hype.
The call: Ask each one: can you solve this with a database? Most could. That’s not a blockchain company; that’s a database company with a token.
The result: The user data agreed. One startup’s own funnel showed a 95% dropout at the wallet-and-KYC step. Same discipline, different hype cycle, applies to AI pitches today. Full story →
Thirty million concurrent connections, on purpose
The situation: ZettaZing needed a push platform that could not fall over at scale, and every off-the-shelf stack had a ceiling below where the business needed to go.
The call: Build ECHO from scratch and own every layer that could break: data format, transport, servers.
The result: In 2014, 3,000+ AWS instances simulated 30 million sustained concurrent connections without a hitch. Designed for billions. Full story →
A bad radio, an unfamiliar accent, a vessel in distress
A Coast Guard crew needed to locate a vessel in distress from VHF audio with a helicopter running behind it. Demo conditions don’t exist out there, so tuning a general-purpose model was never going to be the answer. We engineered for the adversarial environment itself and carried the system through federal authorization, which took longer than the engineering did.
I was in the operations room when the crew used it to find that vessel. That work is now AMBIE. Full story →
The software could hear the radio. It could not key it.
The situation: The voice system transcribed two-way radio audio well enough to matter. It had no way to transmit back, because a phone cannot press a radio’s transmit button, and every radio model puts its accessory connector somewhere different.
The call: Key push-to-talk with an audio tone on its own conductor instead of a DC switch closure, so any device could key a transmitter through a plain headphone output. Keep the board passive, so nothing on it needs power and the transistor is driven by the tone’s own energy. Push everything radio-specific out into swappable tails, so one board serves many radios instead of one board per model.
The result: Send and receive. It extended the life of the project. The review that got it there caught a coupling capacitor that would have put the high-pass corner around 2.8 kHz and filtered out most of the 300 Hz to 3.4 kHz voice band, which is the kind of fault that reads as a bad microphone for a week. Connector pinouts were checked against the manufacturer documentation rather than trusted, and one of them was wrong as written.
The $40M acquisition that closed at $28M
The situation: I watched an acquisition nearly collapse in diligence over IP that had never been protected.
The call: The buyer renegotiated. The company had no leverage left, because the gap was real and discovered late.
The result: A 30% haircut: $40M became $28M. IP and technical diligence before the LOI is cheaper than during. This is the discount my diligence work exists to prevent, on either side of the table. The list I run is the technical due diligence checklist; founders can run the pre-diligence self-audit before a buyer does. For a live deal, CodeProvenance.ai is a fixed-fee preflight: the target scans its own code on its own machine, and I write a five-page memo in 48 to 72 hours. Full story →
That’s the visible fraction. Many of the projects I’ve worked on can’t appear here at all: the NDAs are strict enough that I can’t name the client, the industry, or sometimes even the year. If your problem needs that level of discretion, you’re in the right place.
Operating roles
Abundera (founder & CTO, 2026 to present)
Federated platform of 20+ production surfaces under one identity hub. QR safety family, e-signature ceremony with content-bound integrity, financial-advocacy core, fourteen vertical advocacy sites. Cloudflare edge, passkey-only, Plaid-connected. Built solo, shipped from zero in months. Funded; the raise closed in weeks on running product. Across every site I run, on both Cloudflare accounts, over 17 million requests a month.
AMBIE (founder, 2025 to present)
Acoustic intelligence ASR. Patent-pending recognition that adapts to environment in real time. Currently in development.
Voice AI startup (co-founder & CTO, 2018 to 2024)
Voice AI for federal agencies. Coast Guard and DHS deployments. Real-time radio comms ASR in adversarial acoustic conditions, down to the interface hardware that let the software key a radio.
Barbarians Inc (founder & CEO, 2021 to present)
Software studio for startups. Architecture, technical due diligence, voice AI, hard-systems rescue.
Trivlet (founder, 2026 to present)
High-speed competitive trivia. 10,000+ packs, real-time multiplayer, a ranking system that rewards knowledge and speed.
ZettaZing (CTO, 2014 to 2017)
Built ECHO push-notification platform: 3,000 AWS instances, 30M concurrent connections. Designed for billions.
Einstrong / SmartChannels (CTO, 1998 to 2014)
Long-running CMS and prediction-market platform. Acquired SmartChannels into Einstrong in 2004.
Core Logic Software (CEO & CTO, 1996 to 2004)
Consulting and dot-com-era products. RUM (challenge-response spam solution). CoreBA. Mortar Maker.
MSNBC (SDE, 1995 to 1998)
Built Workbench, the in-house CMS that ran MSNBC during the early consumer-internet era. Left before the dot-com crash, which is a story.
What I get hired for
Production systems that have outgrown the people who built them. Voice AI that works in the demo but breaks in the field. Architecture decisions that need a second pair of eyes before the CFO signs. Acquisition diligence that goes past the slide deck. Firmware and circuit work for teams whose software has run out of road and whose problem is now electrical. Hot-sauce business decisions occasionally (separate venture).
What I won’t take
Junior-rate code work. "Digital transformation" engagements that are really change management. Vendor-paid due diligence. Briefs written to validate a decision already made. Anything where the deliverable is a slide deck and not a working system or a written technical opinion. No retainers, no equity: fixed scope, fixed fee.