Artificial intelligence · 2026
Alfred
Intelligence that shows its sources.
- Disciplines
- AI assistant, Voice interface, Real-time graphics
- Technology
- LLM systems, TypeScript, Three.js, WebGL & GLSL, Next.js
01
Challenge
Give traders a conversational analyst that can reason across market data, macro context, positioning and their own trade history — without inventing a single number, and without acting on their behalf unless they confirm.
02
Architecture
- A tool-calling language model whose tools read the same data registry the terminal's widgets use — a failed lookup returns an honest error, never a guessed value.
- Tools that touch the user's account produce proposals; a trade is only recorded after the user explicitly confirms it.
- Tool usage is surfaced in the conversation, so every answer shows which real sources were consulted.
03
Design
Alfred is represented by an orb: a noise-displaced energy sphere with a fresnel rim, an additive glow shell and orbiting particle layers. Its states — idle, thinking, analyzing, responding — map only to real system behaviour, never to fabricated events.
04
Engineering
- Streaming responses with an audio-reactive orb driven by real playback levels while Alfred speaks.
- An opt-in “Hey Alfred” wake word and spoken responses with a speech queue.
- A Visual Mode in which in-browser hand tracking lets the user grab, move and resize widgets with a pinch gesture.
05
Infrastructure
Server-side model calls with per-user context, persistent orb presence across workspaces, and visual-mode tools that are refused unless the overlay is actually open.
06
Outcome
- An assistant whose answers are traceable to the terminal's real data sources.
- A recognizable visual identity that became the reference for the AI entities on this website.
07
Technology
- LLM systemsReasoning layer, always bounded by scoped tools and real data sources.
- TypeScriptShared contract language across client, server and tooling.
- Three.jsAbstraction over WebGL, loaded only where a scene exists.
- WebGL & GLSLGPU rendering surface for visual systems and game technology.
- Next.jsRendering layer and edge of the application — routes, server components, API handlers.
08
Lessons
An AI interface earns trust by being explicit about what it looked at and what it will not do.
Representing machine state visually only works when the states are real.
Building something with a similar shape?
Start a project