DeckLess

Inspiration Giving a presentation means spending hours building slides — then being locked into a fixed order the moment you go live. If you go off-script, you're lost. We wanted to flip this: what if the slides followed you, not the other way around?

What it does Deckless listens to you speak and automatically shows the right slide in real time. You upload your materials, chat with an AI to refine the narrative, and generate a library of slides. During your talk, saying "revenue" surfaces the revenue slide. Saying "team" brings up the team slide. Same slide pool, different path every time — shaped entirely by what you say.

How we built it

• assistant-ui for the chat interface — one component gave us streaming, file attachments, and markdown
• Railtracks for a two-stage agent pipeline: an Ingredient Agent extracts the key message and narrative hooks, then a Creation Agent turns those into structured slides
• OpenAI GPT-4o powering both agents and the chat conversation
• Web Speech API for continuous browser-native voice transcription matched against per-slide trigger words
• FastAPI + Next.js for the backend and frontend, with pypdf/python-pptx for file ingestion and PPTX export
• Animated stats slides that detect figures in your content and count up numbers live on screen

Challenges we ran into The browser's Web Speech API silently times out every ~60 seconds — we had to build an auto-restart loop so voice navigation never cuts out mid-presentation. Structuring the two-agent Railtracks pipeline so the output of one could be cleanly passed as input to the next took iteration to get right.

Accomplishments that we're proud of The slide graph view — an interactive SVG showing which spoken words route to which slides — makes the whole voice-navigation concept instantly understandable at a glance. The animated stats slides that auto-detect figures in your content and count up numbers feel genuinely polished.

What we learned Pre-built UI frameworks (assistant-ui) and agent frameworks (Railtracks) let you move dramatically faster when you actually use them as designed rather than reimplementing what they already give you.

What's next for FlowDeck Real-time transcription with transcription providers for production-grade accuracy, live data slides that pull fresh metrics from PostHog/Supabase at presentation time, and a better self-improvement loop that learns your vocabulary after each talk.

Built With

• fastapi
• next.js