The Dioscuri Project: Zero Shot Safety Auditing & Compliance

The Dioscuri Project: An agentic safety auditor that reads the manual while it watches the floor. We turn raw video into citeable compliance actions to shield every worker.

Comment

Inspiration

Industrial safety is governed by thousands of pages of dense regulatory text, yet enforced by human eyes that can only be in one place at a time. Every year, accidents occur not because the rules didn't exist, but because there was a "Compliance Gap" between the written regulations and the live factory floor. We were inspired by the myth of the Dioscuri: the twin brothers Castor and Pollux, to create a system where vision and logic act as inseparable pillars to protect workers in real-time.

What it does

The Dioscuri Project is an agentic AI auditor that operates as two integrated engines:

Castor (The Watcher): A high-fidelity vision engine that monitors live factory feeds for physical hazards, ergonomic risks, and equipment misuse.

Pollux (The Scholar): A reasoning engine that holds huge regulatory manuals in its active memory.

When a potential hazard is spotted, the system doesn't just flag it; it audits it. It cites the specific regulation (e.g., §1910.178 for forklifts), determines the severity, and generates an immediate, localized message, such as an automated SMS alert in the worker's native language, to stop an accident before it happens.

How we built it

The project is built on the Gemini 3 Pro architecture, leveraging its

  • Massive one-million-token context window,
  • Long thought signature capability,
  • Powerful multi-modal reasoning engine.

Direct Grounding: We uploaded the authoritative 29 CFR Part 1910 PDF directly into the application, allowing for "Zero-Shot" compliance logic without the need for traditional Retrieval-Augmented Generation or complex, time-consuming model training.

Multimodal Fusion: We utilized Gemini’s video processing capabilities to analyze factory footage frame-by-frame, mapping visual spatial data against the regulatory thresholds found in the text.

Agentic Alerting: We implemented AI to trigger emergency notifications in multiple languages (demonstrated with Turkish) based on the calculated risk level.

Challenges we ran into

The primary challenge was "Token Noise." Raw regulatory documents are filled with hundreds of pages of historical footnotes and preambles that can distract an AI. We had to refine our System Instructions to ensure the "Scholar" engine prioritized enforceable mandates over historical context. Additionally, syncing high-speed factory visuals with dense legal citations required precise prompt engineering to ensure the AI didn't "hallucinate" a rule that wasn't there.

Another minor challenge was limiting our demonstration video to three minutes! The features we were able to create but not demonstrate in our video further speak to the power of the Gemini Platform and include:

  • A conversational interface which allows a safety auditor to query the system about what it "saw" in the video, to help better understand the reasoning that was used.
  • A dynamic update to the regulation database which shows the true "Zero Shot Compliance" capability we have achieved.

Time permitting, we invite the judges to watch a longer "Director's Cut" of our product demonstration video: The Dioscuri Project Demonstration Video - Director's Cut

Accomplishments that we're proud of

We successfully demonstrated a system that can cite exact OSHA subparts in seconds. Most impressively, we achieved a zero-shot workflow: the system was not "trained" on these specific safety rules; it "read" them and applied them to the video instantly. This proves that safety technology can be deployed to any industry (mining, maritime, construction) just by swapping the PDF manual.

What we learned

We learned that the "Context is King." The ability of Gemini 3 to keep one million tokens "front of mind" fundamentally changes how we think about compliance. We no longer need to build specialized databases; we can simply give the AI the law, and it becomes a lawyer.

What's next for The Dioscuri Project: Zero Shot Safety Auditing & Compliance

Our goal is to move from Detection to Prediction.

Edge Deployment: Moving the "Watcher" engine to the camera itself to reduce latency to milliseconds.

Predictive Ergonomics: Analyzing worker movement over time to predict musculoskeletal injuries before they occur.

Global Compliance: Expanding our library to include EU-OSHA and international safety standards, making The Dioscuri Project a global shield for the modern workforce.

Built With

  • google-ai-studio
Share this project:

Updates