Real work. Real AI.
No hype.

Benchmarking Models on Procedural Geometry: An Animated DNA Double Helix Third in a small series testing how AI models handle real builds from a cold spec. Tetris tested game logic. The solar system tested 3D math and interaction. This one tests procedural geometry and continuous animation — and it was built mobile-first, designed for a portrait phone screen before desktop. The brief: an animated 3D DNA double helix you can rotate and zoom, spinning continuously to show its twist. The hard part isn't drawing a spiral — it's the parametric math underneath. Two backbone strands have to wind around a shared axis offset by half a turn, with base-pair "rungs" connecting the correct point on each strand at every step. Get the math right and it instantly reads as DNA; get it wrong and you get a tangle. There's very little room to fake it. Built with Cline in VS Code paired with MiMo-V2.5-Pro via OpenRouter — same setup as the other two. https://www.columbiawork.com/mimo25_dna.html

Benchmarking Models on a Visual Challenge: An Interactive 3D Solar System Tetris tests game logic. This one tests something different — 3D math, rendering, and animation. The brief: build an interactive, animated solar system you can orbit and zoom, with the Sun, all eight planets, Saturn's rings, a moon tracking Earth as Earth circles the Sun, a starfield, and visible orbit paths — correct relative motion throughout, smooth on both desktop and mobile. One spec, handed to the model cold, with the freedom to pick its own rendering approach. Like the Tetris build, this came together with Cline in VS Code paired with MiMo-V2.5-Pro via OpenRouter. The trickier requirements here are the ones that separate a model that understands a 3D scene from one that just spins dots in a circle: nested orbits (the moon has to follow Earth, not the Sun), a draggable camera that doesn't flip or trap itself, and pinch-zoom that works on a phone without scrolling the page. Getting those right in one pass is the real test. https://www.columbiawork.com/mimo25_solar.html

Benchmarking Local + Hosted Models with a Classic: Building Tetris I've been running a small experiment to compare how different AI models handle a real coding task from scratch — not a toy snippet, but a complete, playable HTML5 Tetris with collision, rotation, line clears, scoring, level progression, and touch controls, all specced out in a single build document and handed to each model cold. This build came together using Cline (the open-source agentic coding extension for VS Code) paired with MiMo-V2.5-Pro, Xiaomi's reasoning-and-code model, routed through OpenRouter. The interesting part wasn't just the model — it was the harness. Earlier attempts with a different setup kept stalling on the file-editing step, where the model would write solid code but fumble applying it cleanly. Swapping the agent harness to Cline, same class of task, and it read the spec file itself and wrote the whole game in one pass. The lesson worth keeping: for agentic builds, the harness matters as much as the model behind it. Total cost for the run: about five cents.

Three iterations on the Precoa Portland office wireless migration runbook, each grounded in real data rather than assumption. v1.1 incorporated findings from a full read-only Meraki Dashboard API audit (18 devices, 2 organizations, 8 open alerts, 6 wireless zones uncovered). v1.2 replaced the original ASCII topology with two proper SVG engineering diagrams — a physical/Layer-2 cable-path view and a logical/Layer-3 trust-boundary view. After uploading a sanitized SonicWall tech support report (2.6 MB, 58k lines, secret-scanned before analysis), the design got an honest reframe: deploy guest wireless first as a low-risk Phase 1 (~45 min maintenance window, blast radius limited to visitor Wi-Fi), defer corp wireless and AWS/Azure VPN updates to Phase 2. Discovered along the way: a retired SAN segment on X4 still passing 290M+ packets, three unused legacy wireless zones, and the default public SNMP community string still configured. All findings rolled into post-project hardening backlog.