Holmes - Deduce, Discover, Decide - Home Screen
Project Architecture
Evidence Library ( Upload and Redact across multimodalities)
Command Center - Witness agentic orchestration in real time for transparency
Command Center - Domain agent routing display from the Orchestrator
Knowledge Graph - Red string board style entity-relationship graph for to find the connections usually not visible to the naked eye
Case Events Timeline
Geospatial Analysis of the Case
Multimodal Redaction supporting PDF, Image, Audio and Video files

🕵️ Holmes: AI-Powered Legal & Investigation Intelligence Platform

💡 Inspiration

Legal professionals and investigators spend a staggering amount of their time manually reviewing evidence, cross-referencing documents, videos, audio, and images searching for connections and contradictions. Holmes was born to change that. By leveraging domain-specialized AI agents, Holmes transforms weeks of manual analysis into minutes of transparent, citation-grounded intelligence.

🔍 What It Does

Holmes lets investigators input large volumes of multimodal case evidence (MP3 audio, MP4 video, JPEG/PNG images, and PDFs) then orchestrates a team of specialized AI agents to do the heavy lifting.

🤖 Specialized Agent Team:

Financial Agent - Tracks money trails, transactions, and anomalies
Legal Agent - Identifies statutes, precedents, and violations
Evidence Agent - Evaluates authenticity and cross-references sources
Knowledge Graph Agent - Resolves entities and maps relationships
Synthesis Agent - Detects contradictions, gaps, and cross-domain links
Geospatial Agent - Analyzes location intelligence and movement patterns

🖥️ Five Interactive Views:

Agent Flow - Real-time visualization of the agentic pipeline in action
Knowledge Graph - All case entities and their relationships, like a digital red-string board
Timeline - Chronological event reconstruction
Geospatial Map - Location intelligence and movement tracking
Verdict Dashboard - Case strength assessment at a glance

💬 Smart Chat: A contextual Chat Agent answers questions grounded in clickable source citations, so every claim traces back to the evidence.

📓 Investigator's Notebook: Captures voice and text notes on the fly.

🔒 AI-Powered Redaction: Natural language-driven censorship across all media types:

PDFs → Black box redaction
Images → Blur & pixelate
Audio → Bleep sensitive content
Video → Blur, pixelate, or black box

Gemini identifies targets, then applies pixel-level censorship. For video and image redaction, self-hosted locally deployed instances of SAM2 and SAM3 handle precise segmentation.

🛠️ How We Built It: Gemini 3 at the Core

🧠 Deep Thinking & Reasoning

All agents use Gemini 3's ThinkingConfig at HIGH level via Google ADK's BuiltInPlanner, enabling the Orchestrator to reason about routing and Synthesis to cross-reference findings holistically.

🎨 Native Multimodality

Gemini 3 models process PDFs, videos, audio, and images natively. Domain agents receive raw files with configurable media_resolution (HIGH for forensic analysis, MEDIUM for strategy). Files ≤100MB go inline; larger files use Gemini's File API (up to 2GB).

📚 1M Token Context Window

Stage-isolated ADK sessions feed Synthesis Agents all generated findings, entities, and relationships in one call, enabling cross-domain contradiction detection that's usually not possible with smaller windows.

⚙️ Architecture

9 Gemini 3 agents orchestrated via Google ADK with:

PostgreSQL-backed stage-isolated sessions
Pro-to-Flash fallbacks for resilience
Parallel execution across domain agents
SSE streaming with thinking traces
Tool-based Chat for grounded Q&A

The Orchestrator agent has full autonomy to decide how many instances of each domain agent to spawn based on case requirements derived from initial triage.

🏗️ Tech Stack

Layer	Technologies
Frontend	Next.js 16, React 19, D3.js, React Flow
Backend	FastAPI, Google ADK
Infrastructure	Cloud Run, Cloud SQL, GCS

🌍 Impact

Holmes transforms complex investigation from manual review into AI-augmented intelligence, surfacing connections humans miss, detecting cross-modal contradictions, and generating hypotheses with full provenance. Every conclusion traces to its exact source, building the trust that legal work demands.

🚀 What's Next for Holmes

🔧 Immediate Enhancements

Production Deployment Fixes - The production deployment is currently flaky due to the lack of stress- and chaos-testing, and might be buggy at times.
More Domain Agents - Expand and refine the agent roster with specialized agents for compliance, HR investigations, insurance claims, and more
Multi-Language Support - Enable evidence processing and analysis across multiple languages for international investigations
Google Drive Integration - Allow direct evidence upload from Google Drive for seamless workflow integration

Note: It is highly recommended to test the app locally to experience its full scope. Please send a mail at pluses-snow6q@icloud.com to enquire about the environment variables needed to test the app locally.

🔭 Long-Term Vision

Horizontal Expansion - Create purpose-built editions of Holmes:
- Holmes Legal Research - Specialized for case law analysis, statute interpretation, and legal brief generation
- Holmes OSINT - Tailored for open-source intelligence gathering, social media analysis, and digital footprint investigation
Enterprise Features - Team collaboration with role-based access control, audit trails, advanced reporting dashboards, and organization-wide knowledge sharing
No-Code Agent Builder - Enable non-technical users to create and customize specialized investigation agents through an intuitive drag-and-drop interface