The Tower
Twin-GPU generation rig — images, video, voices. Heavy work happens here, one heavy thing at a time.
patricktanna.com
Patrick Tanna
Building boutique AI systems for work, play, and discovery.
CNN → Sony Pictures → Nordstrom Rack | HauteLook → Central Group → Amazon → Honest Financial Technologies · and after dark, the lab
on the bench
Split keyboard configurator
Eight split boards, part by part: cases in wood and metal, keycap sets, legends, per-key light.
[listening for “hey honey”]
brain
room › [silence]
honey ›
pipeline
- WAKE “hey honey” · own wake word 92% recall
- ENDPOINT VAD + semantic end-of-turn 0.60s
- HEAR streaming scout + verifier · EN · ไทย 0.18s
- ONLY YOU? voice match · optional 0.21s
- THINK · say conversation brain · on The Brain · + memory graph 1.31s first clause
- THINK · do 4B, grammar-locked · can’t invent a tool parallel
- GATE confirm · allowlist · audit · kill policy
- SPEAK local voice · a filler covers the wait 2.48s first audio
- FACE / ACT emotion + lip-sync · lights · TV v→v 5.29s
tail -f logs/actions.log[replayed evening · real medians · no mic, no model]
▍
[avatar · Bear Commander, a Live2D rig by @KenKenMo6 · from Fluffellow on Etsy ]
Honey
Wake word to voice to face, every action through a gate. On-device by default; cloud opt-in, PII scrubbed.
A sword forge’s configurator
A real-time 3D configurator for a Thai sword forge: every blade, wrap and scabbard generated in code, the carry rope simulated, in English and ไทย.
A code-drawn study of the movement seen through the caseback: rose-gold plate with Côtes de Genève, a platinum seconds ring, a jewelled three-armed bridge, and three nested cages turning in eight steps a second around a balance beating at 4 Hz.
A watch that doesn't exist yet
42 mm · 4 Hz · eight days · constant force
[Ayutthaya · 1563 · schematic, not to scale] ✦ wat · ◆ invented household · ··· chapter 1, by boat
Kirin of the Golden City
A story bible for a game: one family, one city, the first fall.
elsewhere
AI-Patrick
A real-time digital twin that runs entirely on one desk.
[real-time avatars · speech & voice · applied AI]
The Archive
Twenty years of one life, parsed into a graph that answers questions.
[knowledge graphs · data engineering]
FangDai · ฟังได้
Live English rooms, captioned in Thai, per speaker, on-device by default — for Thai speakers who are deaf or hard of hearing, living abroad.
[accessibility · speech & voice · native apps]
The Trading Brain
A self-improving research loop that must prove itself on paper — and ask a human — before anything real.
[quant trading · knowledge graphs · self-hosted infra]
Even Glasses
An AI coding agent in a pair of smart glasses — tool permissions approved with a finger ring.
[applied AI · native apps · speech & voice]
Creator Studio
One local app over the whole media-generation stack — image, video, voice, music.
[applied AI · broadcast & media · self-hosted infra]
Honest Analytics
Customer business analytics with automated reporting for a credit-card fintech — every query runs as the viewer, never a service account.
[product leadership · data engineering]
Risk Controls
A 55-tab control-testing spreadsheet, rebuilt as a role-based app — monthly test runs, evidence attached, an append-only audit log.
[product leadership · data engineering · self-hosted infra]
[sounding: 7 strata · − 72 m]
stratigraphic record
Still being deposited
Back in Bangkok, building credit products for Indonesia’s mass market — people holding their first card. I study spending habits the way I once studied shopping carts, except the win is different now: good credit habits taking root, limits earned rather than granted, and the quiet pride of watching your own financial health grow.
[archive: Honest Financial Technologies · 2024 –]
[archive: first-card credit products · Indonesian mass market]
[archive: this layer is still being deposited]
Some habits I coach. Some I keep.
Personalization at planetary scale
At Amazon I tuned personalization for hundreds of millions of people I would never meet. The machine learned tastes at a scale no single mind could hold — and some nights, walking home in the rain, I wondered what it would mean to point all of that at exactly one person.
[archive: Amazon · Senior PM · P3 ML personalization platform]
[archive: +18% Prime acquisition, Q3 2022 YoY]
The org chart said senior. The apartment said barely moved in.
The city where both languages worked
Bangkok was a return, not an arrival — the language was already mine. I ran product for a retail empire in a city that never fully sleeps, and when the pandemic locked the doors, groceries went digital overnight and our systems carried four times the weight without breaking.
[archive: Central Group · Head of Product → VP of Product]
[archive: 4× online grocery surge, 2020 · held]
[archive: EN/TH bilingual]
A city that never fully sleeps will keep any vice you bring it company.
Teaching storefronts to read minds
In Los Angeles I learned that a storefront is a stage. Movie franchises first, then fashion — I taught search engines to guess what a stranger wanted before she finished typing it. One year they gave the company-wide customer-obsession award to exactly one person, and it was me.
[archive: Sony Pictures Interactive → Onestop → Nordstromrack.com | HauteLook]
[archive: Principal PM → Senior Manager, Search & Personalization]
[archive: Customer Obsessed Award, 2018 · sole company-wide recipient]
[archive: guest lecturer, UCLA Extension — agile product management]
I taught machines desire while practicing very little restraint of my own.
The basement of broadcast
Television was a river and we were the riverbed. Eight years at CNN — first keeping the machines alive, then explaining them to the business. I watched the world go from standard definition to high, one rack at a time, and learned that every pipeline is a promise someone has to keep at 3 a.m.
[archive: CNN · 2003–2011 · systems → business analysis]
[archive: SD → HD transition · render farms · media pipelines]
The machines were easier to keep alive than everything else I was supposed to.
Machines first, then the business of machines
I learned the machines first and the business of machines second — computer engineering, then later an MBA to understand why good systems die in bad organizations. Both degrees from the same place. Some roots you keep.
[archive: B.S. Computer Engineering, Georgia Tech]
[archive: M.B.A., Scheller College of Business · marketing]
[archive: certified scrum master, 2013]
I learned systems fast. People took me twenty more years.
Before the feed
In Georgia in the eighties, the feed was a food court. You showed up, you got seen, you learned what everyone was doing by walking a slow lap past the arcade. Presence was the only platform, and logging off meant going home for dinner. I have spent every year since trying to keep what mattered from that world while helping build the one that replaced it.
[archive: photographs in the vault begin 1998]
[archive: messages begin 2006]
[archive: everything earlier lives only in the first narrator]
Not everything from that world survived me. Some of it I let go dark on purpose.
the connective tissue
A map of disciplines and the projects that join them. Move between points with the arrow keys; press a point to hear what it touches, and press it again to let go.
[archive: 13 projects · 13 disciplines · traced on contact]
the civilization
Most of it never leaves the house.
The Brain
A laptop with the memory of a small datacenter. Hosts the large local minds.
The Keeper
Small, silent, always on. Serves the archive and keeps the promises.
The Vault
Twelve terabytes of everything — datasets, weights, twenty-eight years of photographs.
The Mesh
The invisible roads between them. Prompts and tokens travel; the weights stay home.
the underground sky
[archive: export 2026-07-28 · aggregates only · no one is named]
[archive: older layers detected · excavation pending]