dustin hydenengineer + designerontario, canada
nowsenior software engineerside project · d1g, steerable agent workflows

hyden — dustin hyden, engineer and designer

i build language models from scratch, a desktop app that runs a dozen AI agents at once, and the interfaces between people and both.

scroll ↓that's me on the cursor. he's clickable.
(00)about, briefly

i train language models from scratch on a graphics card from 2014. i run a dozen coding agents at once from a desktop app i built for exactly that. and i design the screens people actually touch.

205M
parameters in d2, the model i’m training now
32,768
tokens in the tokenizer i trained for it
47
tokens a second, d2 chatting offline on a gtx 970
4,940
commits into tixert, where i’m cto
01

the model

d → d2 · pytorch · kaggle · one gtx 970

a tiny phi-3-style language model i'm training from zero. it reads about a billion tokens a week on kaggle's free GPUs, then comes home to chat offline on my desktop's gtx 970.

parameters
205M (142M thinking)
shape
20 layers × 768
vocab
32,768 byte-level BPE
context
2,048 → 8,192 tokens
trains on
2× t4, kaggle free tier
training speed
~10,400 tokens/s
chat speed
~47 tokens/s on a gtx 970
tuned with
sft + dpo
tools
calculator · sandbox · apps
d2 v1.3 · a real session, real time · offline on a gtx 970
d · the first one · 49M params, trained start to finish on the 970
how d2 does a word problem
  1. reads the problem
  2. writes a tiny program
  3. a sandbox runs it
  4. six tries vote
  5. the reply must contain the result
(how d2 gets made)hover a step
01 datareading mix v4a billion tokens of general reading first, then 60% general, 30% math, 5% jokes, 5% chat.
02 tokenizer32,768-token BPEtrained on its own mix: 9.5% more text per token than phi-3’s, and every digit is its own token.
03 shape20 × 768205M parameters, plus a 16,384-row bigram table.
04 pretrain2,048 → 8,192every release ends with a long-chat stage, so an 8k chat still fits in 1.4 GB on the 970.
05 tunesft + dporeal conversations, small talk, and better-liked vs less-liked joke pairs.
06 toolscalculator, sandbox, apps“open the calculator” opens it. vague requests get a question back first.
~/lab/context-snakeplayable · 1 of 3

context snake

a sliding context window you can steer. it holds 24 tokens, so every new one pushes the oldest out the back. run into yourself and it hallucinates.

click the screen, then arrows or wasd. space restarts.

tokens seen 0 · context 4/24 · best 0
context: (empty)
02

the orchestra

d1g · tauri · typescript · rust

one window.
every agent.

d1g is the desktop app i live in. claude code, codex, ollama, or a model running on the machine itself, side by side, each in its own folder. i hand work to one agent or a whole group, and it gets done off screen while i keep going.

d1g board with a chat on the left and live cards joined by arrows
01the board. notes, tables, flows and chats are live cards; arrows show what mentions what.
d1g sidebar listing many agents, with a plan table assigned across them
02every agent in one sidebar, working or idle. this plan's eight tasks are split across three of them.
in turn
summarise the routes card
tick the second one
draft the release note
next
fix the failing test
now
answer this first
→ fredy 7 · working off screen
03nothing handed to an agent gets dropped. it waits in line: in turn, next, or now.
a d1g table card made by an agent
04agents use the cards through d1g's own MCP server: find it, read it, change it, report back on it.
d1g/tables/plan-what-is-this.json
"task": "check what you changed by hand",
- "status": "to do",
+ "status": "done",
"who": "fredy 7"
changed by fredy 7showundo
05no database, no server. every card is a plain file with a change log, ready for git. undo one agent without losing another's work.
01 / 05
keep scrolling →

MCP,
without the guesswork.

hand a model a raw API and you've handed it a sequencing problem: it has to remember the right order on every call, forever. so i built yapi, a deterministic layer between agents and the backend. the model says what it wants done; the server decides the order, the same way every time, whether the caller is a model, curl, or a UI. it ships with its own MCP adapter.

POST /yapi
yapi: 1
items:
  - do: account.modify
  - do: order.place
auth→accounts→inventory→billing→notify

every item visits every station in order; stations that don't handle its kind pass it through. the line is the contract, and it lives on the server.

d1g runs an MCP server too, so any agent can find, read and change your notes, lists and tables, then report back on the card the work came from.

MCP servers on this page
  • d1g cards
  • yapi adapter
brains d1g can drive
  • claude code
  • codex
  • ollama
  • lm studio
  • openai-compatible
  • ibm granite, offline
~/lab/herd-the-agentsplayable · 2 of 3

herd the agents

orchestration without the orchestrator. twelve agents wander; you are the scheduler. they move away from your cursor, so steer them into the done lane.

0/12 in the done lane · 0.0s
03

the canvas

seat maps · 3d · motion · react, three.js, canvas 2d
tixert / seat map builderlive product

draw rows. number seats. drag a section. sell it.

the seat map builder i made for tixert. before ai. venues lay out their room, group seats into sections, and give each section its tickets.

(the 3d character)

the little red figure is me.

a low-poly 3D model of me, rigged and animated for my 2023 site. there are two of him on this page: one stands on the “hyden” wordmark at the top and turns to follow your cursor; the other walks with you along the bottom of the screen as you scroll. click either one.

his animations
  • idle
  • run · scrolling down
  • pull · scrolling up
  • clicked
  • knocked out · 3 quick clicks
how it's built
model
character.glb · 1.3 MB
made for
hyden.ca, 2023
texture
a single 16 × 16 colour palette
animations
5, blended over 180 ms
on this page
2 copies, one WebGL renderer
camera
orthographic: 1 unit = 1 screen pixel
resolution
drawn at 42% for the pixel look
before all this

hyden house, my web studio, 2019 to 2025. sites and brands for small businesses.

six of them ↘
suniela beach
on the mend therapy
bee by the sea
manor12
mr. valisier
alta lake electric
04

the box office

tixert · co-founder & cto · 2025–

ticketing built for venues.

contracts, signatures, seat maps, recurring events, landing pages and checkout. soft-launched in canada on march 14, 2025. i'm co-founder and cto: i lead the engineering, the platform, and the tools around it.

4,940commits and counting
2025soft-launched march 14
venue radarin-house · tixert team only

venue radar is our in-house tool for finding venues. it grid-searches a whole city, one search point at a time, and turns what it finds into prospects for our sales. built for us, not for tixert's users.

05

the path

business → ads → web → studio → startup → engineering
2010 → now
201020122014201620182020202220242026
2010 – 2014laurentian universitybachelor of business administration, marketing
sep 2014 – apr 2015omnicom media groupcoordinator, digital ad platforms · toronto
2015 – 2018freelancedeveloper
2018 – 2020georgian collegeinteractive media design, web
2019 – 2025hyden housefounder, web studio
oct 2019 – may 2021city of barrieweb development lead · contract
mar 2025 – nowtixertco-founder & cto
may 2026 – nowh2 analyticssenior software engineer · remote
2014 – nowendless side projectsthe through-line
workmy ownschool
06

index

a few standouts, not the full list · hover a row
#namewhatmade withwhen
01d1gone desktop app for agents, notes and projectstauri · typescript · rust2026
02d2a small language model, trained from scratchpytorch · kaggle2026
03dthe first one. 49M params on a gtx 970pytorch2026
04tixertticketing built for venues. co-founder & ctonext.js · graphql · prisma2025–
05venue radarour in-house tool at tixert for finding venuesweb · mapsin-house tool
06cadencean experiment: project management rebuilt on event sourcingtypescript · event logexperiment
07yapia deterministic layer between AI agents and your backendtypescript · MCP adapterexperiment
08hyden.cathe 2023 site, where the figure was bornreact three fiber2023
09hyden housemy web studio: sites and brands for small businesseswordpress · react · next.js2019–25
10livebarriea live entertainment brand, strategy to identitybrand · web2022
> mounting /hyden .......................... ok
> loading weights: d2 v1.3 .................. ok
> rigging character.glb ..................... ok
> compiling shaders ......................... ok
000%