Tecnologías VM
Work
Hackathon finalist · 2025

Canopy Intelligence

Canopy Intelligence reached the finals of the Financial AI Hackathon Championship 2025 (LandingAI), on the Strategic Investment Timing track. Built by our team Pura Vida Sloth, it maps emerging technologies onto a Gartner-style Hype Cycle by combining knowledge graphs with GraphRAG over multi-source market data.

Two views of one system: the Hype Cycle the agents draw, dissolving into the knowledge graph they read it from — same graph, same chart. The amber dot is the detected contradiction: the crest on the chart, the hub in the graph.
Demo

See the system run.

The Canopy Intelligence demo from the hackathon — the five-phase GraphRAG pipeline and the live Hype-Cycle frontend.
Imagery

From New York to the Hype Cycle.

Real moments from the Financial AI Hackathon Championship in New York, and the Hype-Cycle charts the system produces for the eVTOL industry.

Opening talk at the LandingAI hackathon
Financial AI Hackathon Championship, NYC 2025
Co-founder Oscar Montes with Andrew Ng
Canopy Intelligence Hype Cycle for eVTOL, with the technology maturity timeline
The full knowledge graph and intelligence foundation the chart is read from
01 / 05
14Independent data sources
12LangGraph agents
4Temporal intelligence layers
86.5%Extraction fidelity across 52 SEC filings

Its thesis is cross-layer contradiction analysis. It reads four independent temporal layers: innovation signals, market formation, financial reality and narrative. Their disagreement is the signal worth acting on, so when media saturation peaks while insiders sell and innovation slows, that contradiction surfaces timing 12 to 24 months ahead of consensus.

Under the hood it runs a five-phase, fully reproducible pipeline: 14 data-source collectors, multi-format document processing (using LandingAI's Agentic Document Extraction for high-fidelity SEC extraction), pure-GraphRAG ingestion into Neo4j Aura, a 12-agent LangGraph state machine, and a real-time FastAPI and WebSocket backend driving a D3.js frontend.

Every score traces back to source documents, so analyses are auditable and the same graph always produces the same chart. The project became our deep dive into GraphRAG as an architecture pattern, and into LandingAI's Agentic Document Extraction. Both are now reference techniques we reuse.

How it reads the market

Four layers, read against each other.

The signal is the contradiction between independent layers: surfaced 12 to 24 months ahead of consensus.

18–24 mo lead

Innovation signals

Patents, research papers, GitHub activity.

12–18 mo lead

Market formation

Government contracts, regulatory filings, job postings.

0–6 mo

Financial reality

SEC filings, insider trades, holdings, prices, earnings.

Lagging

Narrative

News (GDELT) and press releases.

The pipeline

Five phases, end to end.

  1. 01

    Phase 1

    Multi-source collection: 14 collectors with checkpoint/resume, rate limiting and parallel downloads.

  2. 02

    Phase 2

    Document processing: multi-format parsing, GPT-4o-mini extraction, LandingAI ADE for SEC filings.

  3. 03

    Phase 3

    Graph ingestion: pure GraphRAG in Neo4j Aura; raw data and relationships only, never derived scores.

  4. 04

    Phase 4

    Intelligence: a 12-agent LangGraph state machine scoring each layer, then Gartner phase detection.

  5. 05

    Phase 5

    Real-time API: FastAPI + WebSocket streaming into an interactive D3.js Hype-Cycle frontend.

Contact

Let’s build something that works.

Have a real problem where emerging technology might be part of the answer? We’d like to hear about it.