i'm alan —i build intelligent systems.

The video is the prompt

A multimodal pipeline feeds live video frames to Gemini, which continuously describes the scene and steers a generative-music model — the soundtrack shifts as the stream does. A text prompt is optional seasoning, not the input.

Live rooms, streamed audio

Viewers join live streams through a Discover/Trending/Library UI; generated audio arrives as WebSocket chunks kept in lock-step with the video, sub-400ms end to end — half the latency I started with.

Won, then spotlighted

Won MLH's Best Use of Gemini API at HackTX 2025, then selected for the Google Gemini Developer Spotlight. Bonus mode: an ambient piano scored by live NOAA space-weather data.

One prompt, many cursors

At the shell prompt every participant edits the same command line; keystrokes merge conflict-free via a CRDT and Enter from anyone runs it.

Tools run only on the host

Claude Code, vim, htop run in a real PTY on the host and stream to every peer, so collaborators never install a thing. Raw/alt-screen modes are auto-detected.

GPU terminal + a live editor

A from-scratch Metal renderer with a CoreText glyph atlas, plus a built-in /edit editor (RGA CRDT, syntax highlighting, remote cursors) and a bundled encrypted tunnel.

Steer the model with a patch, not a prompt box

Prompt, Instrument, Vocals and Output nodes with weighted connections, so you shape the mix instead of retyping prompts.

Go live, then capture

Stream generation in real time from Google Lyria, then grab instrumentals and vocals straight onto a multi-track timeline.

Arrange and export

Clip splitting, volume and pitch automation lanes, and mix export to WAV. A focused desktop tool, shipped as an Electron app with tagged releases for macOS, Windows and Linux.

Record every failure

A plugin wraps the opencode runtime and records every tool call, edit, rejection and postflight check, classifying failures into a graph in Postgres + pgvector.

Warn before it happens again

Before a new task runs, its text is embedded and matched against the failure graph — ANN search plus a 3-hop traversal — injecting a system addendum and, when warranted, hard-blocking tool calls.

One command, zero config

curl | bash brings up the whole stack: Postgres, the FastAPI service, and a Next.js dashboard for exploring the failure graph. Won 2nd overall and the Cognition Company Challenge at LA Hacks 2026.

Whiteboard first, types later

Phase 1 is a freeform collaborative whiteboard. When 80% of the room votes, Claude vision reconstructs the sketch into a typed node graph of real AWS resources.

An agent in the room

Phase 2 pairs the graph with a shared chat where the agent critiques the architecture and edits the graph directly via tool-use, live in front of everyone.

Vote to ship

A second 80% vote compiles the graph with AWS CDK and deploys it to LocalStack. Won Best Use of AWS at Hook 'Em Hacks (UT Austin).

The cursor is data, not pixels

Recording captures the screen and a separate high-frequency cursor/click telemetry stream. Because the cursor is never baked in, zooms are generated from your clicks after the fact and a smoothed synthetic cursor is re-rendered at export.

An editor that starts working before you do

Open a recording and zoom regions are already suggested at each click position, with cinematic easing. Trim, crop, per-segment speed, backgrounds, motion blur — every effect stays editable until export.

Studio in a box, all on-device

DeepFilterNet voice enhancement (~37dB noise reduction at 25× realtime), on-device Whisper captions in 13 languages, one-click silence and filler-word removal, and hardware HEVC/H.264 export via VideoToolbox. No accounts, no uploads, no telemetry.

Push-to-talk, streaming as you speak

Hold Right ⌥ and a floating HUD shows the partial transcript live — WhisperKit runs on the Apple Neural Engine, so nothing leaves the machine and audio is never written to disk.

An LLM tidies what you said

On release, a local Ollama model strips fillers and fixes grammar, punctuation and tone at four intensity levels — from raw passthrough to full polish. The only network traffic is to localhost.

Types into whatever is focused

The cleaned text lands at your cursor via the Accessibility API (clipboard-paste fallback), with voice commands ('scratch that', 'new line'), custom vocabulary and per-app formatting.

MRI in, structured read out

Upload any T1/T2 slice and Gemini Vision returns a diagnostic summary — tumor type, grey-matter loss, abnormalities — in seconds instead of a radiology queue.

Chat grounded in the scan

Follow-up questions are answered against the structured JSON context from the analysis, so the model talks about this scan, not brain MRIs in general.

Every run archived

Each analysis is persisted to S3 and surfaced in a History tab — built, demoed and placed 2nd overall in one weekend at Dartmouth.