-
-
NavourSync AI
Inspiration
During a conversation with an HR representative, I discovered how much time production teams waste manually verifying crew dossiers before every shoot — checking IDs, bank statements, image rights forms, and handwritten contracts one by one. The process is slow, error-prone, and exposed to payroll fraud. That conversation sparked the idea: what if an AI agent could do all of this automatically, in seconds?
What it does
NavourSync AI is an autonomous multi-agent HR compliance system for film and TV production crews. It processes a complete crew dossier in under 10 seconds:
- 00-VISION classifies each uploaded document (ID card, bank RIB, image rights release, employment contract) using Gemini 2.5 Flash multimodal vision
- 00-AUDITOR performs structured OCR extraction, checks expiry dates, and cross-references the bank account holder name against the crew member's ID to detect potential payroll fraud
- 00-DISPATCH generates an HR decision badge —
VALIDATED,PENDING, orBLOCKED— plus a confidential internal note for the production manager and a ready-to-send email in both English and French - 00-ANALYTICS persists every audit result in ClickHouse Cloud in real time, powering 4 live dashboard metrics: total documents processed, average compliance score, expired IDs, and fraud alerts
How we built it
I designed and built NavourSync AI with the help of Bob from IBM, who guided me through the architecture and helped significantly with the code.
The system is built natively on the Google Agent Development Kit (ADK). The three agents are declared as FunctionTool objects and orchestrated by a single OrchestratorAgent:
orchestrator_agent = Agent(
name="navoursync_orchestrator",
model="gemini-2.5-flash",
tools=[
FunctionTool(classify_document),
FunctionTool(audit_document),
FunctionTool(generate_crew_notice),
],
)
AI: Gemini 2.5 Flash via google-genai SDK, with temperature=0.0 on the audit agent for deterministic OCR
Analytics: ClickHouse Cloud via clickhouse-connect — real-time audit insertion and live dashboard queries
Infrastructure: Google Cloud Run, Vertex AI ADC authentication, GCP Secret Manager
UI: Streamlit with a James Bond 007 dark theme
Challenges we ran into
The biggest challenge was authentication. Despite having an active billing account with GCP credits, I kept hitting free-tier quota limits because my API key was the wrong type (AQ. OAuth token instead of AIzaSy...). The solution was to migrate entirely to Vertex AI with Application Default Credentials on Cloud Run — no API key needed at all.
Navigating the Google Cloud Console was also a steep learning curve. Between enabling APIs, creating service accounts, configuring IAM roles, and setting up Secret Manager, there were many moving parts to get right before a single line of agent code could run in production.
Accomplishments we're proud of
Getting the full pipeline to work end-to-end in production — real documents, real Gemini API calls, real ClickHouse writes — with zero mock data. Submitting a working, deployed project to a Google hackathon as an independent developer is something I'm genuinely proud of.
What we learned
The Google Cloud Console is like a cockpit — powerful, but you need time to learn every instrument
Google ADK FunctionTool is the right abstraction for building real multi-agent pipelines
ClickHouse delivers sub-millisecond compliance queries even at scale
temperature=0.0 on Gemini is non-negotiable for legal HR data — determinism over creativity
What's next for NavourSync AI
If selected or awarded, I will make NavourSync AI my flagship project and invest serious time into it — potentially building it into a startup. The problem is real, the market exists, and the technology is now proven. Next steps include deploying the ADK OrchestratorAgent to Vertex AI Agent Engine, adding PDF support for multi-page contracts, and extending the system to handle international productions across 10+ document types.
Log in or sign up for Devpost to join the conversation.