Every problem needs a different thing built.
Geocoding infrastructure that answers in eight milliseconds. A memory app for parents. A platform for athletes. An equity calculator. A painter’s storefront. Five live products with nothing in common, which is the point.
Address to coordinate, faster than anyone will do it.
Healthcare, defense, finance. Nothing leaves the network.
Capture, transcription, and editing that keeps the voice.
Sources that disagree, resolved into something usable.
Three products. Three industries that share nothing.
One is judged in milliseconds. One is judged by whether it made someone cry. One has to make a pile of disagreeing devices tell a single story. All three are live. Go use them.
Addresses in, coordinates out, in eight milliseconds.
Zappy Mappy is a geocoding API built for people whose delivery stops moving when the lookup is slow. Drop-off IQ layers delivery intelligence on top: driver instructions, fraud signals, fewer failed deliveries. For healthcare, defense, and finance, the Sovereign Geocoder runs entirely inside your own network, in a container, touching nothing outside it.
- Latency
- 8ms P99
- Versus Google
- 31× faster
- Uptime
- 99.95%
- Deployment
- Cloud or air-gapped
“Don't send us your data. We send you the API.”
at the 99th percentile
A real model turns that sentence into the fields a dispatch system can actually route on.
Twenty seconds of talking, kept forever.
Your kid says something you’d give anything to keep, and you have about twenty seconds before it’s gone. Tucklet takes just your voice, no typing and no homework, and turns it into a written story worth rereading, with the original recording tucked underneath. A different discipline entirely from the work above: here the constraint is emotional, not technical.
- Input
- Voice only
- Effort
- ~20 seconds
- Kept
- Text plus original audio
- Sharing
- Co-parent accounts
“Tiny stories tucked away for someday.”
She said the moon was following us home. I told her it does that for everyone. She said no — it only follows people it likes.
No typing. Press play and a speech model reads it back.
Every device disagrees about the same night.
An athlete wears several devices and logs into more apps, and none of them tell the same story. Calibrated Athlete pulls training, nutrition, sleep, and recovery into a single record, then lets Cali, the AI coach, say something useful about it. Coaches see the whole squad at once; athletes see themselves against the bar.
- Inputs
- Wearables and apps
- Domains
- Training, sleep, fuel, recovery
- For
- Athletes and coaching staff
- Layer
- Reconciliation before insight
“The model was the easy part. Agreeing on last night was not.”
- Whoop7h 12m
- Oura6h 48m
- Apple Watch7h 31m
- Manual log8h 00m
Reconciled against your own history, then weighted by which device actually stayed on your wrist.
I build for work. I build for fun. Mostly I just build.
By day, last-mile delivery logistics at Nash, at a scale where a few seconds of latency is real money. By night the same problems follow me home — Zappy Mappy exists because delivery geocoding was never quite fast enough, and building one that was seemed easier than continuing to complain about it. The rest started the same way: something was annoying, or someone needed a thing.
Principal Engineer. Last-mile delivery logistics — routing, dispatch, and fleet coordination for companies like Walmart, 7-Eleven, and Loblaws.
Built on nights and weekends, all of them running on the internet right now rather than sitting in a repo.
Chemical and biomolecular engineering. Mostly it taught me how far a good simulation will take you.
Technically all of these are passion projects.
Some just got bigger than others. These two are smaller in scope than the three above and came from exactly the same impulse.
A startup equity calculator that models vesting, strike price, dilution across funding rounds, and what exercising would actually cost you. Open source, because the maths shouldn't be a secret.
A storefront for a Colorado landscape painter. Twenty-odd originals, eight collections organised by the places they were painted, sold-state tracking, shipping. No AI anywhere in it, because it didn't need any.
Most useful problems don’t respect categories.
So this list doesn’t either. It isn’t a service menu — it’s what these products, in unrelated industries, actually turned out to require.
Inference and lookup services held inside a 10ms budget, under load, at P99 rather than on average.
Air-gapped, on-premises deployment where data cannot leave the building. Containerised, no vendor round-trip.
Interfaces people use while distracted, one-handed, in twenty seconds, and come back to for years.
Capture, transcription, and model-side editing that tightens writing without flattening whoever wrote it.
Normalising sources that disagree with each other into one record the rest of a product can be built on.
Dashboards for people who manage groups, and the different view for the individual inside that group.
Fraud and anomaly detection where being wrong is expensive in both directions.
Most of a good AI product is the part that is not the model. Drawing that line is most of the work.
What are you trying to build?
If you have something that needs building.
Say what the problem actually is and you’ll get a straight answer: whether this is a good fit, roughly what it would take, and what to do first.