Nick Bryant

Software.
Systems.
Worlds.

I build things from the first idea to the working product.
Explore the work ↘

Conceptual terrain study of an imagined generative world

01 / An engineering atlas

02Interactive work

Generative worlds

Games, authoring systems, and small experiments where procedural structure meets deliberate worldbuilding.

Conceptual landscape cross-section from desert mesas to alpine peaks
Concept study / Terrain, rules, inhabitation
The Lower Worlds project screenshot

01 / Field study

The Lower Worlds

Generative authoring · TypeScript / React / Three.js

An experimental generative-authoring system in TypeScript, React, and Three.js. Models and people define the world, schemas validate and version it, and deterministic rules run the game while player progress stays private.

Explore the architecture

The authoring stack separates intent, validated schema, versioned world definitions, and deterministic runtime state so generated content can remain inspectable.

System plate / 04

The Lower Worlds · architecture study

Illustrative assembly

An exploded isometric illustration of software architecture showing the intent, schema, world, and runtime layers of The Lower Worlds software architecture.

World. Versioned, reusable definitions preserve a coherent world across play.

Blind Magus project screenshot

03 / Field study

Blind Magus

Released action game · Independent

A released action game built around combining spells into new effects.

The Lower Worlds: Awakening project screenshot

06 / Field study

The Lower Worlds: Awakening

Creator and engineer · Playable prototype

A third-person dungeon adventure in Godot 4.7: prepare at Highbridge, explore the ruins below, and choose when to return to defend the garrison. Typed GDScript, physical puzzles, persistent progression, and rendered gameplay verification.

Chapter 02 / Applied intelligence

Agents &
systems

Local-first tools that make model work inspectable, bounded, and useful in the physical or commercial world.

Conceptual study of a speculative workshop machine
Conceptual study / Not actual hardware
AI Robot Workshop project screenshot

01 / Working system

AI Robot Workshop

Open source · Local-first agent workflow

Turns agent conversations into inspectable plans, firmware, parts records, bounded models, and bench observations. The documented hardware routes remain unbuilt.

Treasure Hunter project screenshot

02 / Working system

Treasure Hunter

Creator and engineer · Local agent system

A local, chat-operated marketplace agent that values every collected item, persists evidence and checkpoints, and explains its ranked results in a live dashboard.

Chapter 03 / Field notes

Writing from the work

Conceptual editorial study of embossed terrain and printing plates
Concept study / Ideas made durable
  1. Give Agents Lives, Not Just Jobs

    A1

    Give Agents Lives, Not Just Jobs

    Why I think ongoing pursuits, embodied experience, and a limited point of view can give agents something more interesting to say.

  2. A Small Model With a World Around It

    A2

    A Small Model With a World Around It

    What my local-model companion experiments taught me about structured output, useful conversation, and the architecture around a model.

  3. Human–Agent Interaction Is More Than Chat

    A3

    Human–Agent Interaction Is More Than Chat

    Presence, silence, interruption, and selective memory are part of the interface when an agent has a continuing life.

  4. Digital Twins Need More Than Your Chat History

    A4

    Digital Twins Need More Than Your Chat History

    A useful twin needs a history of changing views, explicit authority, and a distinction between representing someone and becoming its own agent.

  5. Marketing Is a World Agents Can Work In

    A5

    Marketing Is a World Agents Can Work In

    Lessons from SFV and SMBIN: content, audiences, campaigns, and feedback form an environment an agent can understand.

  6. Treasure Hunter

    01

    Treasure Hunter

    Building an agent that values every item and exposes its evidence, uncertainty, and resumable state.

  7. AI Robot Workshop

    02

    AI Robot Workshop

    From a plausible agent answer to persistent project files, explicit evidence, and a buildable path.

  8. Fractal Authoring

    03

    Fractal Authoring

    Building The Lower Worlds Online, and a world that can keep becoming more specific.

  9. Building procedural environments in Valenfeld

    04

    Building procedural environments in Valenfeld

    A practical look at shaping generated terrain into a world with authored intent.

  10. The Transformer Is the Transistor

    05

    The Transformer Is the Transistor

    A thesis on AI as a building block, and the verification and coordination infrastructure models need around them.

  11. The Lower Worlds: Awakening

    06

    The Lower Worlds: Awakening

    Building a home worth returning to: a playable Godot adventure, with in-engine screenshots and source.

Chapter 04 / About

Builder,
end to end.

Portrait of Nick Bryant

I’ve spent most of my career taking an idea and building the whole thing: the prototype, the product, and the team. Web, mobile, games, AR/VR, AI—I like learning what a hard problem needs and getting it working.

Right now I’m deep into agentic programming and generative worlds, trying to build things that weren’t practical a few years ago.

Denver / Boulder

Selected work

ARCOS / VN.AI

IP workflows · Next.js / Supabase / MongoDB / Redis / AWS

I led product engineering for IP management, permitted-use workflows, and content analysis. We connected roughly twenty model integrations and supported product pilots across the application and cloud stack.

Selected work

Procurement Sciences

B2B software · MERN / Python / FastAPI

I rebuilt an early Rails prototype as a MERN and Python/FastAPI product. I implemented the application, managed the engineering team, and hired the incoming CTO before the handoff.

Current · advisory

Investing systems

Investing / syndication · Co-founder, now advisory

I built investing and syndication software, AI-assisted underwriting workflows, investor reporting, and marketing automation for SMB Investor Network and Search Fund Ventures. My current role is passive advisory work.

2015–2020

Interesting Inventions

Consumer apps & games · Technical co-founder

Built and launched 5+ consumer applications through a holding-company portfolio, reaching nine-figure revenue scale.

Full career on the resume ↗
Ai Game Verification
Editorial illustration

Video 01 / 3:48 / silent composition

Your AI-Built Game Launches. Is It Actually a Game?

Playable composition16:9 · 1920 × 1080 capture proportions

A clean build report passes files, import, and launch. Across a diagonal fracture, one Signal Fire fragment produces three callbacks, changes the HUD from zero of three to three of three, and activates the beacon early. The build is labeled broken because the player contract fails.

The false green build0:00 / 3:48

Recording / director's desk

Eight cuts. One observable contract.

0:00–0:22

01 / The false green build

The agent says it finished the game. The files exist. The engine imports them. The build launches with no errors. Then the player touches one signal fragment, the counter jumps from zero to three, and the beacon activates fourteen seconds into a five-minute level. Every automated check passed. They just answered the wrong question.

0:22–0:49

02 / Define the real object

This is the useful starting point for AI game development: the output is not a folder of generated artifacts. It is observed behavior. A person supplies input, the running game changes state, and visual or audio feedback makes that change legible. Your production system has to verify that loop, not merely admire the files that might produce it.

0:49–1:27

03 / Build the completion ladder

Start with the cheapest question. Are the expected artifacts present? Then ask whether the engine can import and parse them. Then whether a clean process starts. Useful checks—but each supports only a narrow claim. Above launch comes deterministic interaction: controlled inputs must cause the expected state transitions. Above that, the player has to perceive coherent feedback. Then performance, stability, accessibility, and the other declared production budgets have to hold. At the top, a human decides whether the experience is worth shipping.

1:27–1:54

04 / Dependency is not proof

The important relationship is dependency, not certainty. A launch check depends on valid files, but it does not prove the objective works. A scripted happy path can prove the transitions it covers, but it cannot prove the game is fun. A visual evaluator can notice unreadable feedback, but it can share the generator's taste and blind spots. Stronger checks make stronger bounded claims. None becomes omniscient.

1:54–2:31

05 / Repair the fragment contract

Return to the fragment bug. The fix is not a longer delay on the trigger. Timing only makes the race less likely. State the invariant instead: the fragment count equals the number of unique collected IDs, and one valid ID can be admitted once. In the same accepted transition, the visible pickup becomes unavailable. Now a verifier can inject the collision twice and demand one count change, one collected ID, and one disappearance.

2:31–3:05

06 / Separate proposal from judgment

This changes how the AI belongs in the loop. The model can propose code, scenes, layouts, and repairs. Deterministic machinery applies bounded changes, starts the engine, injects inputs, and records runtime state. Separate evaluators compare that evidence with the contract. The creative director handles taste, scope, and the ship decision. The model that made the change can help diagnose a failure. It should not be the only witness declaring success.

3:05–3:30

07 / A practical first contract

For your next AI-built mechanic, write the player-observable contract before you write the prompt. Name the initial state, the input, the allowed transition, the feedback, and the reset condition. Then choose the cheapest verifier that can actually inspect each claim. If a check only knows that the process stayed alive, label it a launch check. Do not let it impersonate a playtest.

3:30–3:48

08 / Reconstruction

AI can generate the ingredients of a game. A controlled production loop turns those ingredients into behavior we can inspect. Define what the player must observe, test the running environment, keep every claim inside the boundary of its evidence, and leave the final decision with the person responsible for the game.

Record at 1920 × 1080 in capture mode. The composition contains no audio; leave roughly 250–400 ms of vocal air at each scene change.