Built in-house · May 2026 release

18 months. One product. One platform.

Memo AI is engineered in-house for Memo Fashion. A focused product built around how this company actually works — not a SaaS reskin, not a ChatGPT wrapper, not outsourced.

Every line of code, every cascade decision, every animation, every edge case — hand-written. No templates. No agency. No off-the-shelf widgets.

From the 9-provider fault-tolerant cascade to the pgvector RAG pipeline, the hands-free voice loop, the in-browser Python interpreter, and the Monday-morning AI digest that lands in directors' inboxes — every detail is designed for Memo Fashion's daily workflow. Refined commit by commit with feedback from the team. Maintained every weekend.

The result is a tool that feels like the polished AI products you'd pay £20–£40 per seat per month for — except it's built specifically for how Memo Fashion works, runs entirely in-house on free tiers, and costs the company nothing extra. The same engine also powers Design Mate (technical-pack drafting) and Deals (forex admin) over a shared API.

18
Months building
v2.0
May 2026 release
~120K
Lines of code
£0
Memo's monthly cost

"Memo Fashion deserves an AI tool that feels like our own — designed for our workflow, branded for our team, and free of subscription fees. So we built one, in-house, and we keep tuning it every weekend."

— Memo Fashion engineering

What's in the box

A non-exhaustive list of what ships in v2.0 — 18 months of focused engineering

Fault-tolerant cascade across 9 providers — Groq, SambaNova, Cerebras, Gemini, OpenRouter, NVIDIA, Tavily, Cloudflare, Anthropic. 40+ keys with automatic <50ms failover.
Hands-free voice loop with VAD silence detection — Whisper STT (~150ms) + Cloudflare MeloTTS, full conversational loop, works in iOS PWA.
RAG knowledge base on pgvector — Gemini text-embedding-004, ~800-token chunking with overlap, top-5 retrieval with cosine ≥0.45 threshold.
Artifacts / Canvas side panel — code rendering, sandbox-isolated HTML preview, copy + download with smart file extensions.
In-browser Python interpreter via Pyodide — CDN-loaded, package auto-install on import, sandboxed, no server cost.
Multi-agent runner with SSE streaming — planner → workers → synthesiser pipeline, abort-on-disconnect to save quota.
MCP-style HTTP tool catalog at /api/mcp — bearer-auth, timing-safe compare, CORS allowlist, rate-limited, 7 read-only tools.
Cron-driven weekly AI digest emailed to directors every Monday — pulls 7 days of activity, AI-generates a 5-section brief.
Memory across chats — auto-extracted facts after every conversation, per-user UI to view, add or delete entries.
Persona overlays — HR, Email, Translator, Coder, Designer, Researcher — layered on the base Memo brand prompt.
Folders + sidebar search highlights — collapsible groups, drag-to-folder via menu, persisted collapsed-state per device.
Tool-calling for Memo data — leave balance, attendance, who's in office, expense status, calendar, directory lookups.
Custom SSE streaming pipeline — no external SDKs, full event types (token / thinking / model_used / done).
Native PDF chat — Gemini 1M-token context reads whole PDFs as inline_data, ~700-page ceiling, full layout preserved.
Image generation — Gemini Nano Banana + Cloudflare FLUX.2 klein 9B/4B + FLUX.1 schnell + Pollinations, with edit verifier.
PDF / Word / Excel export of any chat session, branded with the Memo header.
PWA with safe-area insets, offline page, versioned service worker, install prompt with iOS share-sheet fallback.
Per-user in-memory token-bucket rate limits on STT, TTS, export, KB ingest and multi-agent runs.
Temporary chat mode, persona-aware context isolation, sandboxed HTML preview, sanitised stream chunks.
Supabase Management API integration — migrations applied programmatically, env synced across Vercel envs.
Read the full journey →