Link: buildathon.ai Hosts: AI Fund and DeepLearning.AI What: A multi-stage competition to prototype AI products in one day.
Competition format
5 Core Projects: Complete all five to reach Semifinals.
Speed Challenge (Semifinal): An extra project. First five to finish advance to Finals.
Live Demos (Finals): 10-minute presentations of the original five projects to an expert panel.
Winner selection: Highest total points for technical excellence and innovation. Tie-breaker is fastest semifinal time.
Why it matters
Andrew Ng framed it simply: AI coding assistants make prototyping about 10x faster. The real bottleneck shifts from writing code to rapid product judgment, tighter feedback loops, and disciplined evaluation of agentic systems.
Andrew Ng keynote takeaways
Rapid engineering > raw hustle. Execution speed (many reps) is the best predictor of success. AI coding assistants 10× prototype speed; large legacy codebases see ~50%+ speedups. Move fast & be responsible in sandboxes before shipping.
Lower the cost of proof-of-concepts. Don’t force every POC into production; make POCs so cheap that 18 failures are the toll to find the 2 worth hardening.
Code is a less “sacred” artifact. With AI, architectures become two-way doors. It’s fine to try 3 architectures in a day, or even rebuild next week if needed.
New bottleneck: Product management. When building takes a day, waiting a week for user feedback is the drag. Hone intuition fast (play the product, hallway tests, quick surveys) and use data to train your gut, not just to make a single decision.
Best practices evolving fast.
Agentic workflows need disciplined error analysis. Build end-to-end, run evals, then attribute failures to specific steps to focus fixes (not guesswork).
Where AI coding shines/falters. Great on common frontend/backend tasks; weaker where training data is thin (odd research code, low-level GPU corners).
Underestimated vectors: the voice stack (expect many voice apps) & visual AI for documents (agentic PDF extraction is exploding).
On AGI: the term is marketing-muddy; ignore the hype & keep building useful systems.
Talent & education. Huge AI engineer shortage; everyone—not just SWE—should learn to code with AI. CS fundamentals still matter; curricula must add LLM building blocks (prompting, RAG, agents, evals, MCP/voice, error analysis).
Typical builds completed in ~1 hour each
Visual memory search over screenshots with OCR and visual semantics
Voice to Slides with speaker notes and HTML or PDF export
Slack pulse dashboards for sentiment, burnout risk, alerts, and manager actions
Codebase time machine for repo Q&A, timelines, ownership, complexity, and commit insights
Knowledge-graph builder that extracts entities and relations, then answers with Graph-RAG and sources
Finalists:
1. Two Coders and a Finance Guy
Screenshot Analyzer: Upload images/screenshots, extract both text and visual descriptions, keywords, and enable chat-based retrieval.
Voice to Slideshow: Transcribe audio input to build slide decks, outputting HTML/PDF formats with live speaker notes and previews.
Team Sentiment Analysis Dashboard: Connect to Slack, analyze message sentiment across channels, flag burnout risks, recommend manager actions, and individual team member analytics.
Code Analyzer: Integrate with GitHub repos, analyze commit history, ownership, code complexity trends, and enable Q&A chat about codebase features.
Universal Knowledge Graph Builder: Merge text files and URLs to generate complex visual graphs, enabling exploration, filtering, and concept relationships.
2. Unbound (Solo)
Visual Memory Search: Screenshot search by text/object, showing confidence scores for matches.
Slack Pulse Dashboard: Sentiment visualization from Slack messages, split into positive/neutral, including a simple color-coded graph.
Universal Knowledge Graph Builder: Graphs from sample text files, basic visualization.
Codeline Time Machine: GitHub repo analysis with Q&A chatbot.
Voice-to-Slides: Convert uploaded audio files into PowerPoint presentations.
3. Ken Kids
Screenshot Reader: Search and recognize visual content (e.g., color or object) from screenshots using image recognition APIs.
Voice to Presentation: Create slide decks from recorded voice prompts or uploaded audio files.
Employee Engagement Pulse: Analyze Slack channel sentiment (integration issues in demo due to inactive API).
Codebase Analyzer: GitHub repo timeline, feature ownership, complexity analysis, commit history Q&A.
Knowledge Graph Builder: Combines multiple files, vector store embeddings, Q&A referencing sources.
4. Evelina (Solo)
Visual Memory Search: Upload screenshots, search by objects or text, confidence score + explanation.
Knowledge Graph Explorer: Visualize imported text facts, build graph, add URLs, Q&A on content.
Codebase Time Machine: Analyze repo contributors, timeline, configuration, and commit Q&A.
Voice-to-Slide: Create presentations from voice/file, output HTML-slide.
Slack Sentiment Analysis: Sentiment detection and risk scoring in Slack channels with timestamped analytics.
5. Sacred Gamer (Vibe as a Service)
Knowledge Graph Builder: Global and personal graphs using Neo4j/MongoDB, entity/relationship creation, Q&A, concepts tab.
Codebase Time-Machine: GitHub repo analysis—commit authors/pie chart, timeline, AI Q&A.
Visual Memory Search: Screenshot folder search by text/semantic query with confidence.
Voice to Presentation Generator: Live and high-accuracy audio transcription, image integration, speaker notes.
Employee Engagement Pulse: Slack manager dashboard to analyze team/channel stats and burnout, with summing curves and data visualization.
6. Secret Agent
Screenshot Analyzer: Use S3 bucket, extract structured metadata (app type, description, etc.), enable semantic and natural language search.
Universal Knowledge Graph Builder: Upload docs/URLs, create graph databases in AWS Neptune/Bedrock, visualize and query entities, sources shown.
Slack Pulse Dashboard: Sentiment analysis per channel, burnout risk scoring, actionable insights.
Voice-to-Slide Generator: Audio or recorded input, transcribe with Whisper, generate slides/speaker notes via Claude, HTML export.
Codebase Time Machine: Repo analysis with contributor metadata, commit Q&A, caching for popular projects.
Winners
First Place: Sacred Gamer
Second Place: Two Coders and a Finance Guy
Best Design: Sacred Gamer
Solo Participant Award: Iveina
Neo4j Best Use of Knowledge Graph: Team Async
Speakers and judges
Andrew Ng Managing General Partner, AI Fund. Founder, DeepLearning.AI. Executive Chairman, LandingAI. Adjunct Professor, Stanford University.
Chao Peng. Principal Research Scientist, Trae. Leads Software Engineering Lab on AI agents. PhD, University of Edinburgh.
Dwarak Rajagopal. VP and Head of AI Engineering, Snowflake. Led AI or ML teams at Google, Meta, Uber, Apple, and AMD.
Eli Chen. Principal Technical Architect, AI Fund. Early Netflix and Twitter. Co-founded Credo AI.
Jason Koo. Developer Advocate, Neo4j. Community and education for graph databases and emerging tech.
Michele Catasta. President, Replit. AI researcher on LLMs and AI for code. Stanford alumnus.
Ofer Mendelevitch. Head of Developer Relations, Vectara. Built ML at Yahoo!. Co-founded Syntegra.
Paxton Maeder-York. Venture Partnerships, Anthropic. Founder of Alife. Ex Auris Health surgical robotics.
Tengyu Ma. Chief AI Scientist, MongoDB. Former CEO and Co-founder of Voyage AI. Assistant Professor, Stanford.
Yaad Oren. Global Head of Research and Innovation, SAP. Oversees 5,500 plus employees across 20 plus U.S. offices.
How the competition ran
5 Core Projects: Complete all five by the timeline to qualify for Semifinals.
Speed Challenge: Qualifying participants receive an additional semifinal project. First five to complete it advance to Finals.
Live Demo Presentations: Each finalist presents the five core projects on stage in 10 minutes. Highest points win. If there is a tie, fastest semifinal time wins.
Gratitude
Big thanks to the organizers, judges, speakers, and all teams who shipped multiple working apps in a single day.
More: buildathon.ai
Full day livestream: youtube.com/watch?v=Brz7GaUPEDw