01
Meta Ray-Ban Display
- Inputs
- Gaze · Voice · Micro-gesture
- Output
- Lens overlay · Bone-conduction audio
The first consumer surface where attention is the unit of cost. We build companion apps that respect the gram and the second.
§ Lab/R&D arm/Updated quarterly
What we are building
before clients ask for it.
The lab is the part of the studio that buys its own time. A specification, a reading list, an open invitation. These are the artifacts. We are honest about which surfaces we have shipped and which we are ready to ship for the first time. The currency of an institution is its preparation, not its pretence.
On evidence
We have not yet shipped client work on Ray-Ban Display, Vision Pro, Quest, or Wear OS. What we have are written specifications, a public reading log, and slots reserved for first commissions. The page below is structured around that distinction.
§ 01 Surfaces
Five surfaces
your customers will be on next.
Every few years a new surface joins the room a business must consider. We build for them ahead of demand, so when a client decides to invest, the team has already shipped something.
02
Apple Vision Pro
- Inputs
- Eye-tracking · Hands · Voice
- Output
- Full 6DOF spatial
The high-end of spatial computing. We build the editorial layer. Documents, archives, dashboards. For the firms who will fund the next decade.
03
Meta Quest
- Inputs
- Controllers · Hand-tracking
- Output
- 6DOF spatial
Where developer access is widest and the WebXR pipeline is most mature. Our prototyping workshop.
04
Wear OS
- Inputs
- Wrist tap · Glance
- Output
- 1.4-inch display · Haptic
The companion surface. Every glasses-class device needs a wrist. We build the bridges that make the watch the controller.
05
Voice & ambient AI
- Inputs
- Voice · Context · Sensors
- Output
- Audio · Push · Side-channel
The interface without a screen. On-device LLM ranking that respects the battery, the privacy, and the silence.
§ 02 Briefs in commission
Specifications written.
Slots held for first commission.
Each brief is a fully specified 4-8 week build with architecture, stack, deliverables, and a price set privately with the first commissioning client. The first commission on each surface receives a 30% pilot discount, and the right to be named.
Ray-Ban Display reading companion
Long-form articles streamed to the lens sentence-by-sentence. An on-device LLM ranks which paragraphs deserve eye time and which can be skimmed at glance speed. The thesis: a New Yorker piece read in twenty minutes while walking. Brief is 12 pages, architecture diagrammed, latency budget set.
- Timeline
- 6 weeks
- Slot
- Q3 2026 · open
- Deliverables
- Lens app · Pairing companion · Reader ranking model
StackMeta Spatial SDKOn-device LLM (3B)WebXR fallback
Vision Pro spatial archive
A three-dimensional archive browser. Documents float in the room, sorted by recency on one axis and by attention on another. Pick one up, set it down. The shelf remembers. Brief targets law firms, family offices, and architecture practices first.
- Timeline
- 8 weeks
- Slot
- Q4 2026 · open
- Deliverables
- visionOS app · Document spatialiser · Attention model
StackVision OSSwiftUIRealityKitUSDZ
Wear OS glasses-control
The wrist as a controller for the lens. Three taps to dismiss a notification on the glasses, a long-press to mark a paragraph for later. The watch becomes a quiet way to drive the device on your face. Pairs cleanly with the reader above.
- Timeline
- 4 weeks
- Slot
- Q3 2026 · open
- Deliverables
- Watch app · Glasses companion shim · Gesture spec
StackWear OSJetpack ComposeCompanion API
Agentic procurement
An LLM agent that monitors public RFPs across three jurisdictions, drafts a first response in your voice, and posts it for human approval. The only entry on this page we have actually built. For ourselves, and the one we are most ready to productise.
- Timeline
- 2 weeks
- Slot
- Ongoing · commissioned
- Deliverables
- Crawler · Drafting agent · Approval inbox
StackClaude APIPostgresCronWhatsApp Business
§ 03 Field notes
What we have read,
studied, and re-read.
Preparation is the visible currency of an institution. Below is the public reading log we keep on each surface. Official documentation, conference sessions, research papers, and granted patents. None of it is shipped client work. All of it is real preparation.
Meta Ray-Ban Display
- 01Meta Spatial SDK · official docs + dev forum threads
- 02Meta Connect 2024 · keynote + Display deep-dives
- 03Ray-Ban Display whitepaper · public optical-spec breakdown
- 04On-device LLM inference · LLaMA 3B / Phi-3 deployment notes
Apple Vision Pro
- 01Apple Vision OS Human Interface Guidelines · re-read quarterly
- 02WWDC 2024 · "Design for spatial computing" · "Bring your iPhone app to visionOS" · "Render advanced spatial environments"
- 03RealityKit + USDZ reference · pipeline benchmarks
- 04SwiftUI for spatial · gesture, eye, and depth conventions
Meta Quest / WebXR
- 01Three.js + React Three Fiber · production-grade XR patterns
- 02OpenXR spec · cross-platform input mapping
- 03Quest 3 hand-tracking reference · pinch, point, grab
Web applications · 2026 frontier
- 01React 19 + Vite 7 + Tailwind v4 · production patterns from 130+ shipped sites in our ledger
- 02TanStack Router + Query · async-state architecture and the boundary between routing and data
- 03React Server Components · when to use, when to refuse, what they replace
- 04Web Vitals 2025 · INP, CLS, LCP. The new triad and what each actually measures
- 05Edge runtimes · Vercel Edge, Cloudflare Workers, Deno Deploy. Placement decisions for latency
- 06Better Auth / Clerk / Lucia · modern session design after the JWT decade
AI agents · MCP · multi-agent
- 01Anthropic Model Context Protocol (MCP, Nov 2024) · official spec. Tools, resources, prompts, sampling
- 02MCP server design · typed schemas, idempotency, paginated resources, capability negotiation
- 03Claude Code SDK + agent dispatch · patterns from running 30+ design-expert agents in parallel
- 04Tool use design · parallel execution, error envelopes, retry semantics, partial-failure replay
- 05Multi-agent orchestration · Anthropic patterns, LangGraph, AutoGen, CrewAI. When each is right
- 06Agent memory · file-backed persistent recall, semantic dedup, hash-based collision
- 07Eval discipline · golden-path regression, hallucination probes, prompt-cache hit-rate budgets
- 08Open weights · LLaMA 3.3, Qwen 2.5, Phi-3. Quantisation cost and battery profile
WhatsApp infrastructure
- 01Meta WhatsApp Cloud API · webhooks, template message governance, the 24-hour session window
- 02WATI / Vonage / Twilio · integrator trade-offs and ceiling effects (trial-end discovered the hard way)
- 03whatsmeow (Go library) · direct-device pairing via 8-character phone code. Patched in our toolkit
- 04WhatsApp Flows · native multi-step forms, server-side data fetching, ack semantics
- 05Multi-tenancy + number rotation · soft-ban detection, marketing-conversation rate-limit, opt-in flows
- 06Conversational commerce · catalogues, carts, payments, business profile verification
Cloud, data, infrastructure
- 01Render · Vercel · Fly.io · Railway · Cloudflare. Placement decisions and the audit trail
- 02Postgres 16 + pgvector + RLS · the modern stack's single source of truth
- 03Neon · Turso · PlanetScale · edge-data and the read-replica latency question
- 04Observability · Grafana, Sentry, OpenTelemetry. The three rooms a system must live in
- 05Migration patterns · dual-write, shadow traffic, expand-contract, blue-green
- 06Backups + DR · point-in-time recovery, cross-region replication, restoration drills
Cross-platform research
- 01CHI 2024 · Gaze-and-Voice Multimodal Input papers (3)
- 02NeurIPS 2024 · On-device LLM optimisation track
- 03Public patents · Meta US20240XXXXX (display companion), Apple US2024XXXX (gaze-anchored UI)
- 04Smart-glasses ethics literature · attention economics, ambient capture, peripheral suggestion
§ 05 Open problems
The questions we are
publicly noodling on.
Problems we have not yet solved, listed so that the solving is in public. If one of these is the question your business is also asking, we should probably talk.
- 01
On-device LLM ranking that doesn't burn the battery before lunch.
- 02
Privacy-preserving capture on always-on glasses. What stays on-device, what leaves, who is told.
- 03
Calibrating peripheral attention as a UI input. The difference between glance and stare.
- 04
The grammar of micro-gestures across surfaces. What the same tap means on Vision Pro, on Quest, on Ray-Ban.
- 05
Documenting a system fast enough that a junior engineer can ship the second version without consulting the first.
- 06
The ethics of glasses-class commerce. When a brand can suggest a product into your peripheral vision, what should it not do.
§ 04 Workshop output
What we have actually built
and run, every week.
Internal tools the studio depends on. Each is real, each operates in production for our own work, and each is available to be productised for a client commission. This is the infrastructure builder's receipt.
Tool 01In production
Multi-agent design dispatch
The pattern we built to ship 33 bespoke React/Vite websites in parallel during a single fortnight. Each agent receives a per-business brief, scaffolds a fresh project, hand-builds 8-13 pages, audits its own Phosphor imports, verifies imagery context, and reports back. Powers the entire Agri Show campaign on /work.
Claude Code SDKViteReact 19
Tool 02In production
Agentic outreach orchestrator
Personalised email + WhatsApp send pipeline with dry-run gate, IMAP-append for sent-folder fidelity, random stagger, per-recipient hook personalisation, and idempotency on top of an append-only JSON log. The Agri Show 30-email + 18-WhatsApp burst moved through it without an Excel sheet in sight.
Gmail SMTPHostinger IMAPPython
Tool 03In production
whatsapp-cli (patched fork)
A fork of whatsmeow's CLI extended with 8-character phone-number device pairing and a structured JSON-everywhere output mode. The studio uses it for direct-device WhatsApp from the office number when a third-party integrator is the wrong tool. Open to client commission.
GowhatsmeowWhatsApp
Tool 04In production
Prerender + sitemap pipeline
Custom Puppeteer postbuild crawler that visits every SPA route and writes static HTML. Powers SEO for this site's 174 routes. Pairs with a sitemap generator that reads slugs straight from source files (no build-time config drift). Both ship as a single 'npm run build:full' step.
PuppeteerViteNode
Tool 05In production
Aesthetic-preset engine
Hash-based design-token generator that maps one brand brief to a unique aesthetic. Palette, typography pairing, hero variant, signature gesture. Across the 130-site portfolio without two ever looking the same. Eight aesthetic archetypes, hand-mapped by industry then varied per business.
React 19Tailwind v4Aesthetic taxonomy
Tool 06In production
Persistent memory file-system
How we run multi-week campaigns without losing context between sessions. Append-only file-backed memory keyed by topic, with semantic recall, hash-dedup, and explicit overwrite-on-correction. Operates the studio's own agent loop and is being abstracted into an MCP server.
MCPClaudeJSON-LD
§ Lab · invitation
An experiment you think
we should be running?
The lab takes commissions from clients who want a question answered before they fund a programme. Three slots a quarter. Any continent.