Depilot

Inspiration

DePilot was inspired by the friction in real-estate operations, especially in markets where property verification, survey validation, payment confirmation, and deal progression still happen across chats, spreadsheets, screenshots, and manual review. A single land transaction can require cadastral checks, payment reconciliation, buyer verification, and human approval. DePilot was built to make that workflow faster, safer, and more auditable.

What it does

DePilot is an agentic real-estate intelligence system. It accepts natural language requests and uploaded evidence, classifies the task, creates or resumes a workflow, and routes the work to specialist agents.

The LandGuard agent handles survey and land due diligence: coordinate conversion, boundary checks, cadastral risk, missing beacon recovery, and flood/topography analysis.

The DealPilot agent handles deal operations: deal lookup, payment evidence extraction, duplicate reference checks, payment schedule generation, and guarded payment processing.

How we built it

We built DePilot with Google ADK for agent orchestration, Gemini models for reasoning and multimodal extraction, and Firestore for persistent workflow state, memory, sessions, deal records, and audit trails.

The system runs as a TypeScript/Express service packaged for Google Cloud Run. The master orchestrator uses ADK callbacks to prime context, extract deal numbers, classify workflows, create or resume Firestore-backed state, and block unsafe tool execution. Specialist sub-agents then execute typed tools for cadastral analysis, deal verification, and payment workflows.

Challenges we ran into

The hardest challenge was making the system truly agentic without letting the LLM become too free-form. Real-estate and financial operations require autonomy, but they also require deterministic safety gates.

We also had to handle messy real-world inputs: incomplete messages, image-only submissions, payment screenshots, survey coordinates, and requests that mix land verification with deal operations. Designing a workflow classifier and persistent state model helped the agent stay grounded across turns.

Accomplishments that we're proud of

We are proud that DePilot is not just a chatbot interface. It uses a real multi-agent architecture, ADK tool execution, Firestore-backed memory, workflow persistence, and guardrails for sensitive operations.

We are also proud of the domain-specific depth: DePilot combines cadastral intelligence, Nigerian coordinate-system handling, flood/topography review, payment verification, and deal lifecycle automation in one agentic workflow.

What we learned

We learned that useful agents need more than prompts. They need state, memory, tool boundaries, structured outputs, retries, and clear rules for when to stop.

We also learned that LLMs are strongest when paired with deterministic systems. Gemini is effective for classification, reasoning, and extraction, while Firestore workflows and ADK callbacks enforce operational safety.

What's next for Depilot

Next, we want to complete the Pub/Sub background workflow path so DePilot can process asynchronous events without a user actively waiting in chat.

We also plan to add a judge-facing demo dataset, stronger evaluation suites, OpenAPI documentation, production Secret Manager integration, and deeper integrations with official land registries, payment providers, CRM systems, and document verification APIs.

Built With

• firebase
• gloud-pubsub
• google-adk
• google-genai