Portfolio / 2026AI usage research
AI usage research.
I study how AI is actually used: agent workflows, browser games and models running on hardware I own. The focus is the gap between a demo and something that holds up.
- Agent systems
- Browser 3D
- Local inference
- Product experiments
Selected experiments
All experimentsThese are not game-development portfolio pieces. Each one is a technical probe: how far can an LLM carry a real production pipeline — generated assets, code, UI, audio, verification and deployment — and where does a human still have to intervene?
Open world · Xianxia
青云渡 Qingyun Crossing / 修仙开放世界
A 3D cultivation world in a browser tab: eight regions, sword flight, tribulation and a seven-chapter main line.
ModelFable 5.2 + Opus 5.5 · three.js
Cinematic VN · Browser
零号信标 ECHO ZERO / 蓝光游戏
A future-noir visual novel about memory, signal and a city that forgets.
ModelDeepSeek Flash + MiniMax H3 + added voice
Tactics · Browser 3D
烽火三国 Beacon of the Three Kingdoms
A 12 × 9 tactics slice with 47 generals, weather, fire and a complete battle loop.
ModelDeepSeek Flash + MiniMax H3
FPS · Harnessed
BLACK ICE 白夜站 / Arctic FPS
An arctic outpost FPS with four enemy roles, a boss, relays and a complete shell.
ModelPhotoreal-game skill · 8/8 gates

雾谷农庄
DeepSeek Flash · one prompt

拾穗 · Harvest
Blender + DeepSeek Flash · 39 props

机场大亨 3D
DeepSeek Flash + DSH · Godot

NEON BREACH
DeepSeek V4.1 Flash · pure API

落日环线
Sundown Circuit · 3 laps

SUNSET DISTRICT
4 laps · 636 draw calls

明宫三维
1.6-bit · original · Fable 5.2

仙宗 · Immortal Sect
DeepSeek Flash · 7 disciples
Photoreal Game Harness
Eight gates from first light to a shipped 3D build.
Chinese Novel to Game
Ten stages from a novel TXT to an evidence-checked build.
Coding & tooling
All experimentsThe other half of the experiment: tools that make model work inspectable, local-first and repeatable. Built with the same loop as the games, without the game.
MiniDSH
A Python reimplementation of DSH's layers: tools, approvals, sessions, subagents, skills, goals and PTC. MockModel by default.
ModelNot recorded · Python 3.9+ · zero core deps
Open the lab Desktop app / ElectronNorMark
A macOS Markdown reader and editor with CodeMirror, live preview, workspaces, PDF export and focus modes.
ModelNot recorded
Read the build Tooling / Rustsensorium
A terminal hardware monitor: CPU, GPU, memory, fans, power and disks read straight from /proc, /sys and NVML, with no child processes.
ModelDeepSeek Flash · 4–6% of one core
Read the build Spatial tool / Three.jsRoomTessera 3D
Floor plan to editable 3D apartment: wall recognition, furniture, materials, measurement and local saves.
ModelDeepSeek Flash
Read the build Tooling / TypeScriptdsh-quota
Route-aware quota and spend state for DeepSeek Harness: plan remaining, API balance or a local ledger, selected by model route.
ModelDeepSeek V4.1 Flash
Read the buildResearch areas
Full profileAI systems & agents
Agent workflows, RAG, tool use, evals and local inference.
Product experiments
TypeScript, React, Node, Python and static-first deployment.
Spatial interfaces
Three.js, WebGL, Godot and game UI that carries information.
Local infrastructure
Linux, CUDA, Docker and local GPU workflows.
Research & writing
Benchmarks, data analysis and technical notes with caveats.
Interface design
Typography, hierarchy and interaction, built in code.
Local environment
LocalTwo machines, one local-first workflow
A portable Mac for making, a used-parts workstation for local inference, and a deliberately small model bench. The hardware and runtime are part of the research setup, not a procurement flex.
- MacBook Pro
- Apple silicon · 48 GB unified · 1 TB
- Workstation
- 2 × Xeon E5-2696 v4 · 2 × RTX 2080 Ti (22 GB each) · 94 GB DDR4
- Runtime
- vLLM · Qwen3.8-27B-FP8 · 256K context · 180 W cap
- Good at
- Embeddings, classification, local drafts and small LoRA runs.
2 × Xeon E5-2696 v4 · 44C / 88T2 × RTX 2080 Ti · 22 GBNVLink NV2 · 2 × 25.78 GB/s94 GB DDR4
Open source
All contributionsStrata / 43k+ ★
Multi-GPU measurements on old cards
Five upstream PRs on layer split vs one card, batch concurrency and vLLM, measured on 2× 2080 Ti 22 GB over a dual-socket Xeon.
Read the contributionProject Maya / 2.8k+ ★
Client discovery + a 128K report
Exposed context and output limits in /v1/models; reported Maya-XS DeepSeek-V4.1-Flash at 128K on the same pair.
Competitions
All resultsWriting
All notesWriting is AI-drafted · Sources and review: Norwin Yu