1. Inspiration
AI agents are revolutionizing software engineering, but in production, they hit a critical barrier: the context problem. AI agents frequently generate hallucinated queries, broken test cases, or invalid code because they lack visibility into enterprise data ownership, schemas, lineage, and data quality.
We were inspired by Build with DataHub: The Agent Hackathon to ask: What if AI agents didn't just guess context, but were actively grounded by DataHub’s Metadata Graph? We built ADIP (Autonomous Demo & Intelligence Platform) to create a self-governing multi-agent system where every action—from UI exploration to 3D video demo creation and Playwright testing—is backed by real-time DataHub context.
2. What it does
ADIP is an autonomous 8-agent ecosystem that transforms codebase updates and web UI exploration into polished documentation, interactive 3D video keynotes, and automated Playwright regression tests.
Autonomous UI Exploration: The Explorer Agent navigates web applications using Playwright to extract DOM elements, screenshots, and user flows. Knowledge Graph Construction: The Knowledge Graph Agent extracts semantic entities and relationships to build a contextual memory graph. DataHub Context Governance: Every storage tier (adip.raw_exploration_data, adip.knowledge_graph_store, adip.test_results_store) is cataloged as a DataHub dataset URN. Value Generation: The Demo Agent generates text-to-speech narration and WebGL 3D video demos, while the QA Agent runs automated smoke tests and emits data quality assertion passes to DataHub. Live DataHub Context Studio UI: An interactive glassmorphism dashboard (datahub_studio.html) displays dataset lineage, quality assertion scores, and real-time Model Context Protocol (MCP) server logs.
3. How we built it
We architected ADIP across 4 distinct functional tiers integrated with DataHub:
Layer 1: Orchestration: Powered by an Orchestrator Agent that plans tasks, routes requests, and aggregates results. Layer 2: Knowledge Acquisition: Explorer Agent, Knowledge Graph Agent, and Documentation Agent capture UI metadata and generate user guides. Layer 3: Value Generation: Demo Agent, QA Agent, and Release Intelligence Agent render 3D WebGL keynotes and Playwright regression suites. Layer 4: Outputs & Integrations: Exports MP4/WebM videos, PDF/HTML guides, and triggers CI/CD, Slack, and Jira workflows. DataHub Integration: We built a FastMCP Python server (agents/datahub_context/datahub_agent.py) implementing the Model Context Protocol (MCP). It exposes tools like get_dataset_schema, get_dataset_lineage, search_datahub_catalog, and publish_quality_assertion directly to the LLM agents.
4. Challenges we ran into
Audio-Visual Pacing & Synchronization: Syncing AI-generated text-to-speech audio with WebGL camera paths and VTT subtitle timestamps required building dynamic audio duration handlers. Cross-Agent Lineage Tracking: Mapping non-traditional AI data structures (like Playwright DOM snapshots and vector embeddings) into standard DataHub dataset URNs (urn:li:dataset:adip.*) required custom schema definitions. Real-time MCP Tool Protocol: Ensuring low-latency tool calls between Python FastMCP agents and frontend visual state updates.
5. Accomplishments that we're proud of
100% DataHub Metadata Alignment: Successfully mapped all 6 persistent storage engines into DataHub's metadata graph with full upstream and downstream lineage. Interactive 5-in-1 UI Showcase: Created a unified glassmorphism viewer (index.html) featuring the live DataHub Context Hub, Demosmith UI, ADIP App, Director Studio, and 3D Amazon Keynote. Automated Quality Assertion Emitter: QA Agent automatically pushes Playwright assertion results (PASSED/FAILED) into DataHub to ensure zero data drift. Judge-Ready Deliverables Kit: Created an examples/ directory containing complete JSON schemas, lineage graphs, and FastMCP tool call traces.
6. What's next for ADIP + DataHub
DataHub Actions Webhooks: Triggering automatic Demo Agent video regenerations whenever a upstream DataHub schema change is detected. Production DataHub GMS Integration: Deploying the FastMCP server directly onto live DataHub Cloud / Acryl Data infrastructure. Multi-Repo Lineage Graphing: Expanding the Knowledge Graph Agent to trace dataset lineage across microservices and distributed frontend apps. 1:47 AM
Log in or sign up for Devpost to join the conversation.