Drew Davis
I build the platforms that teams ship on at scale—and then go home and build JARVIS, game engines, and real-time AI video pipelines. 17 years of turning “that’s impossible” into “oh, it shipped.”
Building Platforms at Scale
// where infrastructure meets impact
I build the platforms that engineering teams ship on. At Abridge AI, I grew the platform engineering organization from 1 to 25, standing up the Internal Developer Platform (Builders Portal), the cloud infrastructure, Kubernetes clusters, and CI/CD pipelines that let a healthcare AI company move fast without breaking compliance—FedRAMP, multi-region, the works.
The real differentiator is the breadth behind the depth. I built a JARVIS-inspired voice AI assistant called Gambit—Electron + Go + Python, 455 tests, wake word detection to personality-driven speech. I built an autonomous AI agent (Grym) that manages real infrastructure via SSH and ships code through GitHub PRs. I built a real-time AI video engine in Rust that does segmentation at <16ms per frame.
The thesis is simple: curiosity isn’t a hobby, it’s an engineering superpower. The person who debugs GStreamer pipelines at night is the person who intuits why your production system is misbehaving at 3am.
Featured Projects
// what ships after hours
Project Gambit — JARVIS-Inspired Personal AI Assistant
Three-language AI assistant spanning two machines (Colossus desktop + homelab server). Electron shell with a Go sidecar on :9300 (HTTP endpoints, WebSocket bridge, SSE push) and a Python voice runtime. Full voice pipeline: wake word detection (OpenWakeWord) → STT (Whisper) → LLM (Claude with a tunable personality system) → TTS (Kokoro, am_onyx voice) — end-to-end verified. Svelte 5 widget dashboard with a StatusOrb reflecting 5 pipeline states, voice activity visualization, system status, and personality controls (cajun_frequency, humor_level, verbosity sliders with live LLM guideline preview). Personality system: Remy LeBeau identity with tunable Cajun dialect frequency, humor level, and formality. Integration gateway handles Docker container management, system monitoring, and media stack awareness via MCP-compatible tool protocol. 455 tests across the stack (330 voice + 104 gateway + 21 Go). 100+ PRs, 6 gap analysis cycles, solo-built.
Grym — Autonomous AI Infrastructure Agent
Personal deployment of the Hermes Agent framework running on a homelab (Ubuntu 24.04, Ryzen 7 7700, ZFS pool). Discord presence across 10+ channels — each scoped to a domain (homelab ops, project work, client repos, ops log). SSH access to the homelab host, GitHub API integration, autonomous PR creation (branch → commit → PR — never merges its own). 150+ skills covering SSH bootstrap, media stack troubleshooting, cron job management, and more. GitOps model: merge-to-main auto-deploys via self-hosted GitHub Actions runners. Manages real infrastructure: Docker compose stacks, ZFS datasets, Caddy reverse proxy, AdGuard DNS, Cloudflare DDNS, Tailscale mesh. Has built and shipped code across 5+ repos autonomously (gambit, sysw, homelab, null-latency, grym-tech).
sysw — Stuff You Should Watch
Self-hosted media companion app. SvelteKit 5 frontend + Go backend + PostgreSQL + Claude AI recommendations. Arr stack integration (Radarr + Sonarr) with a bidirectional sync engine: FullSync reconciliation, webhook processing, and TVDB→TMDB crosswalk. AI-powered recommendations where Claude analyzes your collection and viewing patterns and explains WHY it recommends something by referencing titles in your actual library. Per-user isolation, privacy-first architecture — AI opt-in defaults to OFF, AI-generated content visibly badged. 7 PRs shipped for arr integration alone. Features include SourceBadge, Send to Radarr/Sonarr action, Ready to Watch lane, scheduled poll sync, and per-user quality profiles. Open-source after beta.
Forge Studio / Project Null-Latency — Real-time AI Video Effects Engine
Real-time AI video effects engine and virtual production studio for Linux. Rust backend manages GStreamer pipelines, ONNX Runtime inference, and virtual camera output via dual-backend support (v4l2loopback + PipeWire). AI segmentation runs in a Python sidecar (MediaPipe Selfie Segmenter) over a Unix domain socket. Svelte 5 frontend provides two modes: a camera studio (NVIDIA Broadcast-style background replacement, blur, effects) and a scene editor (OBS-style compositing). Outputs to virtual camera, MJPEG preview stream (:9002), WebRTC (:9001), and MKV recording. Zero-copy frame pipeline with a latency target of <16ms per frame.
Pixel Bridge — D&D Character-Driven Multiplayer Game Engine
Multiplayer game engine where real D&D Beyond and Pathbuilder 2e character sheets drive gameplay. Go server handles WebSocket connections, a real-time world tick loop, combat system, spatial hash, 54-ability resolver, buff/debuff system, mob AI with pathfinding, and death/respawn — 122 passing tests. Character ingestion pipeline parses actual D&D Beyond and Pathbuilder 2e exports and translates stats, abilities, and modifiers into game mechanics. Unreal Engine 5 C++ game client. Wire protocol: WebSocket + Protobuf. Persistence: pure Go SQLite.
Builders Portal — Internal Developer Platform
Solo-built Internal Developer Platform at Abridge AI. Go/Fiber backend, React/Vite frontend, Cloud SQL. 187 merged PRs in ~4 months as sole primary contributor. 16 backend plugins, 24 frontend pages, full RBAC, and immutable audit logging. 102 MCP tools auto-generated from the OpenAPI spec — AI agents connect via Private Service Connect tunnel, fully private, no public internet. Features: Terraform run visualization with cost impact, push-button deployment wizard with dependency graphs, ArgoCD dashboard with sync triggers, cloud asset inventory, image catalog with SBOM integration, and a gamification leaderboard turning drift remediation into a competitive sport. Replaced manual Terraform workflows with self-service for all of engineering.
What I Work With
// the toolkit, abridged
Cloud & Platform Engineering
Designing the foundations teams build on
Systems Programming
Close to the metal, far from the crashes
AI/ML & Computer Vision
Teaching machines to see, hear, and think
Game Dev & Real-time
Where latency budgets are measured in frames
Full-Stack Web
From database to deploy button
Infrastructure & Homelab
Production-grade infra, closet-grade budget
Data & Optimization
Making numbers tell the truth, faster
Compliance & Security
Regulated environments, hardened systems
Beyond the Day Job
// where curiosity meets compilers
Voice AI from Scratch
Built a complete voice pipeline: wake word detection → speech-to-text → LLM reasoning → text-to-speech. Not using a framework — raw audio processing, WebSocket bridges between Go and Python, and a personality system that makes Claude speak with a Cajun accent.
Real-time at 60fps
Game servers that synchronize state across clients at tick rate. Spatial hashing for collision, pathfinding for mob AI, ability resolution for 54 D&D abilities. When your latency budget is one frame, you learn distributed systems from the inside out.
AI That Ships Code
Grym doesn't just answer questions — it SSHs into the homelab, clones repos, writes code, runs tests, and opens PRs. 150+ skills, channel-scoped permissions, destructive operation guardrails. A production AI agent, not a chatbot.
Zero-Copy Video Pipelines
Rust + GStreamer + ONNX Runtime. Real-time AI segmentation at <16ms per frame. Dual virtual camera backends (v4l2loopback + PipeWire). The kind of systems programming that makes you dangerous in any stack.
The Homelab as a Platform
Ubuntu 24.04, ZFS, Docker Compose, Caddy with wildcard certs, AdGuard DNS, Tailscale mesh, GitOps deploys via GitHub Actions. Same disciplines as cloud platform work, but the on-call rotation is just me at 11pm.
Let's Talk
// open channels
Whether it’s platform engineering, a systems architecture deep-dive, or you just want to see what happens when you ask Gambit to speak Cajun—I’m always up for a good conversation.