Notes from the build.
Hackathon wins, agentic engineering, edge inference, and what we've learned shipping AI into production.
Jack Dorsey's claim — every company can now be a mini-AGI — sounds like a deck slogan until you've actually run an org as parallel agent teams. From inside that loop, here's what changes about hiring, headcount, and what an 'engineer' even is.
Most pieces about AI in the editor are about productivity. This isn't. With Claude Code, Codex, Gemini CLI, and Coco running in parallel, AISOFT shipped eight products to real users in a quarter. The change is bigger than 'engineers are faster.' It's that the company shape itself is different.
Studio Copilot won the Raw to Curated track at NVIDIA GTC's Hack to Create. The product is simple: a local AI workspace for photographers that handles culling, contracts, and client review without ever uploading a file. Here's why edge AI for creative professionals is genuinely the right architecture.
At the Nebius hackathon the day before NVIDIA GTC, four of us built Sideline: one Cosmos Reason 2 inference loop driving a SO-101 arm via LeRobot, a MentorPi rover on ROS2, and a Unitree G1 humanoid. Same brain, three bodies. Here's what worked, what didn't, and why physical AI just stopped being a buzzword.
I put NVIDIA's Cosmos Reason 2 on a Jetson and pointed it at amateur tennis. The model doesn't just see frames — it reasons about ball trajectory, line calls, and the physics of a moment. Here's what that buys, and where it falls down.
Two MacBook Pros, two operating systems, three flash tools, an OTA upgrade that bricked the bootloader, raw PyUSB debugging, and one kernel parameter change to flash an NVIDIA Jetson AGX Orin 64GB. Over 20 attempts across two days. Everything that went wrong and the one-line fix that finally worked.
Most AI side-projects don't make it past stage two. Here's the funnel I now run every product through — what each stage is for, what kills you between them, and the gate that lets you move forward.
What changes when the same product runs on a $2,000 box under your desk instead of a per-token API. Latency, cost curves, what breaks, and the LiteLLM-as-router pattern that lets you flip between local and cloud without rewriting agent code.
One weekend, one Jetson, 2.3 million Austin construction permits, and a Nemotron Nano 8B running locally. How a small team beat larger ones by leaning all the way into agentic engineering.