Verifying the physical location of an asset, whether a factory, a shipping container, a real estate development, or a supply chain node, from an image is a complex intelligence task. It requires synthesising visual cues with semantic reasoning.
The Challenge
1. The Challenge: Visual Verification and OSINT
Verifying the physical location of an asset, whether a factory, a shipping container, or a supply chain node, from an image is a complex intelligence task. It requires synthesising visual cues (landmarks, vegetation, signage, architecture) with semantic reasoning.
It is not enough to recognise “a building”; the system must recognise “a Haussmann-style building, likely in Paris, consistent with a late afternoon sun angle.”
Technical Architecture
2. Technical Architecture: The Graph-Vision Hybrid
We led the end-to-end development of a Geo-Semantic Classification Pipeline that integrated disparate AI modalities into a cohesive reasoning engine.
2.1 The Knowledge Graph (Neo4j)
We constructed a knowledge graph to codify relationships (e.g., “Eiffel Tower” → located_in → “Paris”). An LLM was fine-tuned on Cypher code to allow analysts to query the graph using natural language.
2.2 Multi-Modal Vision Pipeline
- Vector Database (ChromaDB): Images were embedded to enable visual similarity search (k-NN).
- Object Detection & OCR: Extracted text (road signs) and cross-referenced them against the knowledge graph.
- Physics Sanity Checks: A specialised module analysed shadow angles to infer time of day and rough latitude. If the model predicted “London” but the shadows indicated a tropical latitude at noon, the prediction was flagged.
2.3 Reasoning Engine
The Phi-3 Vision model served as the reasoning core. It integrated Image + OCR + Shadow Data to output a reasoning chain. A final Meta-Scorer (MLP) weighted inputs dynamically based on confidence scores.
Operational Impact & ROI
- Speed: The entire pipeline processed jobs in ~5 seconds via multiprocessing and multithreading, delivering intelligence-grade localisation at scale.
- Reliability: The inclusion of physics-based checks (shadows) and structured knowledge (graph) significantly reduced “hallucinations” common in pure generative vision models.
This case study describes work undertaken by the founder of Altablack prior to the firm’s creation, presented here to illustrate the technical and strategic foundations of the practice.
