01
Stack
Next.js, FastAPI, PostgreSQL, Celery and Redis, WorkOS SSO on Azure AD, OpenRouter for text and image models, Recall.ai for meeting capture, AWS ECS, RDS, and ElastiCache.
Outcome
One place where a project's meetings, tasks, brief, and designs live and can be asked about. Designers start from a reviewed brief and rendered directions instead of a blank page, and human review sits at the brief stage so nothing reaches a client without a designer's sign-off.
AI design proposals from client meetings
- Context
- A design-and-build firm that fits out retail stores for consumer brands. Every project starts with client meetings, and every proposal starts with a designer re-reading notes to reconstruct what the client asked for.
- Problem
- Requirements lived in meeting recordings and memory. Turning them into a brief, then into storylines and moodboards, took days of designer time before any creative work began.
- What we built
- An internal AI platform with a shared shell for single-sign-on, navigation, and tool discovery, and a set of tools on top of it that follow a project from first meeting to design.
-
Meeting intelligence
A bot joins Microsoft Teams calls, transcribes them, and extracts a structured record: summary, action items, decisions, and project facts.
-
Task tracking
Action items from meetings become tracked tasks with owners, so follow-ups stop living in chat threads.
-
Design generator
Transcripts become a design brief mapped against the firm's own questionnaire, with gaps and confidence flagged per field for a designer to review. From the brief the system generates one to five creative storylines with narrative, colour and material palettes, and AI-rendered moodboards, storefronts, and interiors. Designers regenerate with feedback, compare storylines side by side, and roll multiple meetings into one brief.
-
Floor plan to design
Starting from a store's floor plan, the platform produces design directions for the actual space rather than a generic concept.
-
Grounded assistant
A question-answering layer over the firm's meetings, briefs, and designs, so an answer about a project cites the meeting or document it came from instead of guessing.