Grant Harris

Problem

A local model needs more than a prompt box. You have to see what it’s doing to the hardware, keep long conversations inside the context window, and keep an agent away from anything it shouldn’t touch.

Built

Streaming chat with reasoning traces and six per-turn metrics; live GPU, VRAM and memory telemetry; automatic context-window management; and an agent mode with 14 sandboxed tools locked to one workspace folder, with a diff to approve before each write in Ask mode.

How I checked

A test suite, plus a benchmark script that measures whether the “auto” reasoning mode actually saves time on a given machine.

Python · JavaScript · Bash

Code↗

Built

A 1-billion-parameter model (MiniCPM on llama.cpp) routes tasks to specialist prompt bundles through a SQLite-backed sidecar.

Results

A frozen suite of 24 tasks, averages per task

Setup	Tier	Passed	Time	Tokens
Small model alone	GPU	16/24	3.4 s	1,589
Small model + router	GPU	17/24	0.8 s	282
4× larger model	GPU	24/24	1.7 s	341
Small model alone	CPU	17/24	16.2 s	1,265
Small model + router	CPU	14/24	6.4 s	285

How I checked

A frozen task suite and repeatable A/B runs on two hardware tiers. On GPU, the router beat the same model alone on quality, speed and tokens. On CPU-only hardware it lost on quality, and a 4× larger model passed everything. The write-up says so.

llama.cpp · SQLite · Python · Model evaluation

Code coming soonResults coming soon

Built

A pipeline that scores a game on each of the 27 ingredients, using a fine-tuned model that reads only a text profile of the game.

First result

On 50 hand-labeled games, the model barely beat predicting the average: it missed by 1.37–1.39 points against 1.43–1.46. Across 245 games its scores tracked reviews only weakly, and the hand labels themselves showed no link to reviews, with an average |ρ| of 0.10 where chance gives 0.11.

How I checked

5-fold cross-validation, so every labeled game is scored by a model that never saw it. The first result pointed at the text and the labels, not just the model. A second round, designed but not yet run, uses neutral text, about 200 hand labels, and 30 games labeled by two people to measure agreement before training.

What it became

The 27 ingredients are the foundation of StoryGen, which shapes the story prompts it writes for AI agents with them, and of the AI-guided game design and creation process in Lifestone.

Python · Fine-tuning · Cross-validation · Statistics

Code coming soon

Built

One tap per part of the day on a kiosk app on a repurposed tablet that wakes on camera motion, a server that stays on the home network, and a dashboard. Weather is logged alongside each day, because it can’t be backfilled later. In daily use within a week of the first commit.

How I checked

On the physical device: a crash-recovery routine for the watchdog service and an offline-first queue for dropped connections. Missed check-ins are recorded on purpose, because bad days go unlogged and the data would otherwise flatter them.

Kotlin · Jetpack Compose · FastAPI · SQLite · React

Code coming soon

Built

If hearing tests suggest OTOF-related deafness then prompt genetic testing If gene-therapy eligibility is confirmed then notify the provider If an implant is scheduled without a result then ask the surgeon to confirm Designed around CDS Hooks and SMART on FHIR.

How I checked

A review of 10 sources, requirement tests, an evaluation plan, an FMEA, a stakeholder analysis and a cost-benefit analysis.

Clinical decision support · Research · Health informatics

Paper coming soonPresentation video coming soon

Built

Each run takes a weighted role, reads a shared handoff document, completes one bounded task, and records its work for the next run. EngineerReviewerDesignerPlaytesterManager

Result

Over 300 commits in ten days on a multiplayer game platform, and a fine-tuning dataset built from the agents’ logged failures: reports dumped instead of work finished, runs that hit the turn cap, fumbled tool calls.

How I checked

Review and playtesting are roles in the rotation, so a later run with a different job checks what an earlier run did.

LM Studio · Cron · AI agents · Prompt engineering

Code coming soon

Problem

Keeping a large real-time codebase honest while several AI agents change it at once.

Built

Agents work in parallel lanes with an auditor beside them, each landing small commits with the proof written in the message. Game design and creation are AI-guided, built on the ingredients from Scene Signal.

How I checked

A feature is done when it can be re-proven on demand. Each new check is made to fail once on purpose before it’s trusted, and the main branch only ever moves to a proven commit.

three.js · SpacetimeDB · TypeScript · AI agents

lifestone.io↗