Engineering notes

• A real engine, not a tutorial sandbox: physics, destruction, dynamic lighting and fluid simulation, all driven through no-code authoring
• 2D and 3D on one platform, so a project can grow in scope without being rebuilt somewhere else
• In-house workshops for assets, logic and scenes. The entire creation pipeline lives inside the product, which is the actual moat
• Two 3D engines integrated behind one editor, Three.js and Babylon.js, rather than betting the platform on a single renderer
• Tutorial → Sandbox → Publish → Share, so a complete beginner ends their first session with something playable and shareable by link
• Clerk-backed auth with S3 asset storage over presigned upload URLs
• QA run as a written bug report, every issue tracked with root cause, fix and verification status

Engineering notes

• JustTheMic. The streamer publishes their audio as separate tracks (mic, game, music, guests) and each viewer mixes their own balance in the browser. Turn the game down, mute the music, keep the voice. Nobody else lets the audience control the mix.
• Live captions, translated per viewer. The OBS plugin encodes captions locally in real time, the platform relays them, and the extension translates into whatever language that particular viewer speaks. A streamer broadcasting in one language reaches an audience in all of them, without doing anything differently.
• A native plugin, a cloud relay and a browser extension forming one low-latency pipeline: three distribution channels, three release processes, one product
• 37 numbered development phases with separate public, technical and internal changelogs. The release process is a system, not an afterthought
• Hand-rolled Postgres migration runner with drift history, so schema changes are reversible and auditable
• Real-time Twitch IRC ingestion via tmi.js plus a Discord voice bridge, both surviving reconnects and rate limits
• Encrypted credential storage using @noble/ciphers for third-party OAuth tokens

Engineering notes

• Built for and deployed at a real company. It was part of the TutorMax stack that took monthly tooling spend from ~$15,000 to ~$500
• Removes per-seat licensing entirely. The cost of adding the 200th user is the same as the 20th
• Self-hosted deployment means internal conversations never leave infrastructure the company owns
• Firebase Auth as the single authentication gate, verified server-side with the Admin SDK
• Daily.co voice rooms with lifecycle managed server-side, so clients never hold API credentials
• Ships an Android app alongside the web client, so the team is reachable off a desk without adding a second per-seat bill
• Originally built on Matrix. The federated protocol was ripped out entirely once it became clear the product didn't need it. Reducing scope beat defending the sunk cost.

Engineering notes

• Havok physics simulation running in the browser with a server authoritative over race state
• Two control paths into one simulation: chat-driven AI agents and live WebSocket input from human drivers
• pnpm workspace monorepo splitting game client, API server and shared types
• Custom track and texture pipeline feeding the Babylon scene

Engineering notes

• One codebase, two extension platforms: Chromium and Firefox builds produced from a single esbuild pipeline
• LiveKit WebRTC client running inside a content script for low-latency listen-along audio
• Caption and overlay rendering injected into a third-party DOM it doesn't control, without breaking the host page
• Broadcaster-gated features: viewers only see what the streamer has enabled

Engineering notes

• C++ against the OBS plugin API, a different discipline from the web stack, and proof the toolchain isn't a limit
• CMake presets plus GitHub Actions producing reproducible cross-platform builds
• Ships as a real installer, not a zip of files and a README of instructions

Engineering notes

• Playable browser prototype built first, specifically to answer the design questions before any engine commitment
• A production plan that sequences work so the game is playable at the end of every phase
• Written design documentation covering the core loop, economy and the decisions that are expensive to reverse

Engineering notes

• Region-aware checkout: which payment providers appear depends on where the buyer is
• Full EN/TR internationalisation across catalog, checkout and transactional flows
• Multi-currency pricing with international shipping rules
• One portal, two audiences: customer order history and admin CRM behind the same shell

Engineering notes

• Two-sided marketplace with distinct requester and fulfiller flows over shared data
• Campaign state machine with admin adjudication at the verification and payout steps
• Creator verification pipeline gating who can accept paid work

Engineering notes

• Google Cloud Vision OCR plus a local Tesseract path for extracting totals from receipt photographs
• Shift tracking feeding earnings analytics: hourly rate, best shifts, trend over time
• Free and premium tiers billed through Stripe
• Built from the frontline's actual workflow. The people using it are mid-shift, one-handed, in a hurry

Engineering notes

• ~$14,500/month removed from operating expenses, around $174,000 a year, by replacing rented tools with software the company owned outright
• Meetings became tickets automatically. The AI pulled weekly client-call transcripts through Fathom's API, read what was actually discussed, identified the issues raised, and opened tickets for tech support without anyone taking notes.
• The knowledge base wrote itself. The same pass extracted fixes and context that came up in conversation and were never documented, and filed them into the KB, solving the problem where the answer exists but only in someone's memory.
• Trained on in-house documentation, so its output matched how the company actually described its own systems
• An in-house AI system wired into every department, because all the systems were company-owned, one model could read across meetings, chat, tickets, calls, sales and accounting at once. After a two-to-three month learning period it answered C-level questions on demand and surfaced recurring problems management never heard about.
• The model escalated on its own: a complaint mentioned once was pinned, a complaint that recurred was raised, checked against the internal tooling for an existing fix, and opened as a ticket where none existed
• Dual-audience Kanban: clients watched their own tickets progress, commented, and closed them when an issue resolved itself, while internal technical discussion stayed hidden and the CEO kept oversight of both views
• Separate internal and external knowledge bases from one content model, so a fix written once served both the tech team and the customer
• Hostname-based routing: portal.*, help.* and the apex rendered entirely different route sets and auth regimes from a single codebase and deployment

Engineering notes

• pnpm workspace monorepo on Node 24 splitting the React client from an Express 5 API
• Voice control for step navigation during cooking
• Pantry state diffed against recipe requirements to generate shopping lists automatically