Skip to content

Latest commit

 

History

38 Commits

Folders and files

NameName
Last commit message
Last commit date
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 

Repository files navigation

🛡️ IArgos

Intelligent Investigation Data Analysis System

IArgos is a prototype of an AI-assisted investigation workspace designed to help analysts and investigators organize, categorize, search, and prioritize large collections of images based on an investigation context.

The project explores how multimodal AI (vision + language) can augment human decision-making in sensitive analytical workflows, rather than replace it.

⚠️ This repository is a research / prototype project. It is not production-ready and makes explicit design trade-offs for clarity, experimentation, and explainability.


🎯 Problem Addressed

In many investigative or analytical contexts (law enforcement, journalism, compliance, intelligence, auditing), analysts face:

  • Hundreds of heterogeneous images
  • Weak or missing metadata
  • High cognitive load when deciding what matters
  • Manual, error-prone triage processes

IArgos proposes a structured workflow where AI assists with:

  • First-pass categorization
  • Context-aware relevance scoring
  • Visual exploration and filtering

while keeping the human fully in control of final judgment.


🧠 Core Capabilities

1. Context-Aware Analysis

  • The user defines an investigation context (free-text description).
  • All downstream scoring and prioritization explicitly reference this context.

2. Image Understanding (Vision + Language)

  • Automatic image captioning and visual question answering
  • Extraction of descriptive text and semantic tags

3. Automatic Categorization

  • Each image is assigned to one or more high-level semantic categories (e.g. people, vehicles, documents, weapons, indoor/outdoor, etc.)
  • Explicit conflict resolution logic (e.g. indoor vs outdoor)
  • Transparent rule-based post-processing on top of model outputs

4. Semantic Search

  • Textual search over generated descriptions
  • Synonym-aware and multilingual-friendly queries

5. Investigation Relevance Scoring

  • Custom relevance score (0–100) combining:

    • Detected categories
    • Richness of description
    • Semantic overlap with investigation context
    • Exact and partial keyword matches
  • Images classified into three operational buckets:

    • 🟢 Relevant
    • 🟡 To review
    • 🔴 Likely irrelevant

6. Analyst-Centric UI

  • Multi-page Gradio interface
  • Clickable statistics and filters
  • Progressive disclosure: overview → drill-down

🏗️ System Design Philosophy

This project intentionally avoids an end-to-end opaque model.

Design choices emphasize:

  • Interpretability over raw accuracy
  • Deterministic scoring logic layered on top of AI outputs
  • Explicit heuristics that can be inspected, modified, or debated

This makes IArgos suitable as:

  • A research prototype
  • A discussion artifact with domain experts
  • A foundation for more robust, audited systems

🧪 Technical Stack

  • Language: Python

  • UI: Gradio (multi-page Blocks)

  • Vision-Language Model: BLIP (captioning + VQA)

  • Architecture:

    • Stateful session-based analysis
    • No persistence (in-memory only)
    • Modular scoring and categorization pipeline

🔬 What This Project Is (and Is Not)

It is:

  • A serious exploration of AI-assisted analytical workflows
  • A demonstration of multimodal reasoning pipelines
  • A portfolio project showcasing system-level thinking

It is not:

  • A production-grade forensic tool
  • A claim of automated decision-making
  • A benchmark-optimized ML system

📌 Current Limitations

  • Heuristic-based scoring (not learned end-to-end)
  • No long-term storage or audit logging
  • Performance not optimized for large-scale datasets
  • Model bias and error propagation not formally evaluated

These limitations are explicit and intentional at this stage.


🚀 Possible Extensions

  • Replace heuristic scoring with learned ranking models
  • Add embedding-based semantic similarity (instead of string matching)
  • Introduce active learning from analyst feedback
  • Formal evaluation on real investigative datasets
  • Stronger security, logging, and access controls

👤 Author Intent

This repository is part of a broader effort to explore:

  • Human-centered AI
  • AI as a decision-support system
  • Practical multimodal pipelines beyond demos

It is shared for discussion, critique, and learning, not as a finished product.


⚖️ License

MIT License


If you are a researcher or practitioner interested in investigative AI, explainable systems, or applied multimodal pipelines, this project is meant to be read, questioned, and challenged.

About

Vision-language triage for forensic analysts: BLIP-2 + VQA to classify and sort images recovered from seized devices, replacing manual review of large media dumps.

Topics

Resources

Stars

0 stars

Watchers

0 watching

Forks

Releases

Packages

Contributors

Languages