Inspiration
In highly regulated global industries such as cosmetics and chemical formulations (governed by EU Regulation (EC) No 1223/2009), bringing a formulation to market requires assembling an exhaustive Product Information File (PIF) and safety dossier. Today, this process creates severe enterprise friction:
- Formulation scientists, toxicologists, and regulatory assessors spend hundreds of manual hours extracting parameters from heterogeneous supplier Safety Data Sheets (SDSs) across PDFs, Word documents, spreadsheets, and CSVs.
- Mathematical safety margins must satisfy the EU SCCS 12th Notes of Guidance, where a single calculation error can trigger product bans or regulatory recalls.
- The Enterprise Adoption Barrier: Enterprises already have massive investments in Product Lifecycle Management (PLM), LIMS, or ERP systems (SAP, Veeva Vault, Oracle). They cannot and will not replace their entire tech stack for AI compliance.
We were inspired to build FortifiedReg Fleet: an Embeddable, Component-Based Multi-Agent Governance Engine on Google Cloud Run. Instead of a monolithic replacement, it is designed as a pluggable regulatory intelligence sidecar (Hexagonal Ports & Adapters) that embeds directly into existing enterprise workflows, combining Gemini 3.7 / 3.6 Flash, Google Model Armor guardrails, deterministic toxicology calculations, and cryptographic Human-in-the-Loop (HitL) verification.
What it does
FortifiedReg Fleet operates as an embeddable regulatory component engine providing autonomous compliance microservices:
- 5-Format Supplier Document Intake: Autonomously parses and normalizes supplier technical files across 5 binary formats (PDF, DOCX, CSV, XLSX, PPTX) with OOXML container validation and SHA-256 digital fingerprinting.
- Deterministic SCCS Toxicology Calculator: Calculates Systemic Exposure Dosage ((SED)) and Margin of Safety ((MoS)) with zero hallucinations: $$SED = \frac{A \times \frac{C(\%)}{100} \times \frac{DA_p(\%)}{100} \times F_{ret}}{BW}$$ $$MoS = \frac{NOAEL}{SED} \ge 100 \quad (\text{Certified Safe})$$
- INCI Restriction Cross-Referencing: Instantly verifies ingredients against EU CosIng Annex II (prohibited substances) and Annex V (preservative thresholds).
- Google Model Armor Defense: Real-time screening blocking prompt injections, path traversals, and malicious inputs before reaching internal agents.
- Human-in-the-Loop (HitL) Checkpoints: Locks decisions in a tamper-evident state machine with lease fencing, requiring Product Manager cryptographic sign-off before generating certified PIF dossiers.
- Interactive Gemini Regulatory Copilot: Live multilingual consultation backed by Gemini 3.7 / 3.6 Flash for formulation optimization and IFRA standard citations.
🧪 Demonstration Harness & Simulation Sandbox: The public web portal (
https://fortifiedreg-fleet-251114662133.us-central1.run.app/) is an interactive test harness designed specifically for hackathon evaluation and auditing. It simulates real-world dual-role enterprise governance (R&D Formulator & Product Manager) and provides live feature sandboxes demonstrating the underlying component APIs.
How we built it
FortifiedReg Fleet is engineered on Hexagonal Architecture (Ports & Adapters), enabling its domain kernels to be embedded into any host framework:
- Component Decoupling: Pure domain calculators and intake ports are completely decoupled from persistence and delivery layers, allowing easy integration as REST microservices, gRPC services, or background worker sidecars.
- Serverless Compute: Deployed on Google Cloud Run (
us-central1) with serverless autoscaling and scale-to-zero configuration for zero idle cloud costs. - AI & Agent Intelligence: Powered by Google Gemini 3.7 / 3.6 Flash via the Google Agent Development Kit (ADK) / GenAI SDK for structured toxicology extraction and regulatory grounding.
- Security & Guardrails: Integrated Google Model Armor for prompt injection scanning, combined with stateless HMAC-signed JWTs enforcing strict server-side Role-Based Access Control (RBAC).
- Secrets Management: Google Cloud Secret Manager securely provisions and rotates HMAC signing keys (
fleet-jwt-secret) with IAM service account accessor binding. - Supply Chain Integrity: Immutable, digest-pinned OCI container images built via Google Cloud Build and hosted in Google Artifact Registry.
- Auditability & Observability: Immutable OpenTelemetry-style audit sink with deterministic UUIDv5 event deduplication and sanitized Google Cloud Logging.
- Official Google GenAI SDK & ADK Integration: Built with the official Google GenAI SDK (
google-genai) (from google import genai) and Google Agent Development Kit (ADK) underpackages/fleet-adapter-google-adkandapps/fleet-api/src/fleet_api/routers/assistant.pyfor autonomous toxicology extraction and live regulatory copilot reasoning powered by Gemini 3.7 Flash / Gemini 3.6 Flash.
Challenges we ran into
- Designing for Zero-Disruption Enterprise Embedding: Ensuring that the core governance engine can slot cleanly into existing enterprise IT topologies without imposing database schema migrations or breaking host authentication mechanisms.
- Fail-Closed Governance in Asynchronous Workflows: Designing multi-agent state machines that refuse to fail open. We implemented strict lease fencing and transactional sagas ensuring that when unexpected errors or role escalation attempts occur, the system safely halts with sanitized error codes and zero credential leaks.
- 5-Format Binary Document Normalization: Handling edge cases across diverse binary file formats, such as complex Word XML style compression, malformed PDF text layers, and nested spreadsheet sheets, while preventing zip-bomb decompression attacks.
- Strict Zero-Leak Logging: Ensuring that internal file paths, raw tenant payloads, and authentication tokens are comprehensively redacted before entering public API responses or Cloud Logging streams.
Accomplishments that we're proud of
- True Modular Embeddability: Proved that pure regulatory domain logic can be encapsulated into decoupled, testable components that integrate seamlessly without rebuilding an enterprise tech stack.
- 312 Automated Tests (100% Pass Rate): Comprehensive test suite covering domain calculators, adapter boundaries, lease concurrency, and 22 fail-closed negative security gates.
- 11/11 Live Remote Probes Verified: Automated remote verification CLI (
verify_remote.py) confirming 100% compliance directly against the live Google Cloud Run production deployment. - Zero-Setup Interactive Web Portal: An interactive web portal allowing judges and users to experience a complete dual-role compliance walkthrough and 6 interactive sandboxes with zero login friction.
- Production-Ready Packaging: Fully packaged with one-click automated deployment scripts (
deploy.sh/deploy.ps1) and zero-cost teardown automation (destroy.sh).
What we learned
Enterprise AI adoption hinges on modularity and governance. High-stakes compliance workflows do not need monolithic "all-in-one" black-box LLMs that disrupt existing ERP/PLM investments. Instead, building componentized, deterministic, and verifiable multi-agent kernels with Google Cloud serverless security allows enterprises to adopt agentic workflows safely, incrementally, and immediately.
What's next for FortifiedReg Fleet
- Pre-Built ERP/PLM Connectors: Out-of-the-box integration plugins for Veeva Vault Quality, SAP S/4HANA PLM, and Oracle Agile PLM.
- Global Jurisdiction Expansion: Extending the regulatory engine from EU (EC) No 1223/2009 to US FDA MoCRA, ASEAN Cosmetic Directive, and China NMPA compliance frameworks.
- Continuous Post-Market Vigilance Agents: Proactive AI sidecars monitoring global toxicology literature, CosIng updates, and SCCS opinion drafts to automatically alert brand owners of emerging formulation risks.
Log in or sign up for Devpost to join the conversation.