A survey of work and hobby projects, 2024 to 2026

Gavin Daher

I take problems from my hobbies, try the model that might solve them, and measure the answer against real data.

Stations
7 hobby, 4 from the day job
Architectures
17 tried
Longest run
9 days, autonomous
Day job
Founder, Physical AI Team at Denali; stations D1 to D4, ending at Summit

hobby stationits traverse
day-job stationits traverse
Every object is rendered under the same light. Textured objects are real outputs, stone-coloured ones are measurements, and grey card ones are stand-in models, not photos.

Contours: the Manaslu Himal, Nepal (28 to 29°N, 84.15 to 84.97°E), every 400 m. AW3D30 ©JAXA.

Cast of a learned cave network, rendered in stoneThe learned cave in a real Minecraft client, as a mounted tile Station 1. Research, August 2026Neural KarstogenesisCaves learned from 12 real ones Bar sculpture of experiments run per day, the busiest day in orange Station 2. Agent harnessNine days autonomous564 runs; 13 to 17 August by day Minotaur generated from one concept imageOne-eyed creature, auto-riggedSkull guitar pedal generated from pixel artRetro computer generated from an illustrationCarved stone face generated from an illustration Station 3. Generative 3DAssetForgeOne image in, a rigged model out, all local Terrain from a fine-tuned diffusion model, as a mounted relief tile Station 4. Research, 2026Terrain diffusionA LoRA, then two programs closed by their rules Pumpkin character from a FLUX style LoRA, as a mounted tileSpace-station market from a FLUX style LoRA, as a mounted tile Station 5. Fine-tuningStyle LoRAs and synthetic data4 diffusion models tuned, plus persona LoRAs Orbital Atlas world map, as a mounted tile Station 6. Agent teamOrbital AtlasA map mod built by agents, 9 weeks to beta Bar sculpture of the agent ledger: runs, leases, parked questions, and human decisions in orange Station 7. HomelabThe lab92 human decisions in 17,672 events

Day-job traverseFour stations from my work at Denali Advanced Integration, 2025 to 2026. These are my employer's systems, so each grey object is a stand-in model, not a photo of the real system.

Stand-in model, not a photo of the real system: a returns tote holding a folded garment, a box and a bottle, with a camera on an arm above it Station D1. Day job, 2025Denali ReLoReturned goods sorted by a vision-language model Stand-in model, not a photo of the real system: a packing bench with a depth camera on a post, a marked measuring area with three items, and an open carton Station D2. Day job, 2026Pack Bench of the FutureMeasures an order, recommends the smallest safe carton Stand-in model, not a photo of a real part: a flange with a cracked spot and a blue box drawn around the crack Station D3. Day job, 2026Rainier-VLA 2.1 B-parameter inspection model for edge GPUs Stand-in model, not a photo of the real system: blocks on a track passing one blue gate, with a mark on the blocks that have passed Station D4. Day job, 2025 to 2026SummitThe harness where Claude runs my team's ML work