storeops

6-customer-counter
5-mongodb-aggregation-engine
3-grounded-answer-receipt
1-off-duty-owner-console
2-evidence-trail-mongodb
4-review-to-action-atlas-vector-search

Inspiration

Walk into any small shop and you'll find one person doing ten jobs at once. The moment the owner steps away to serve a customer or clock out, orders go unanswered, shelves go unwatched, and reviews go unread. This hackathon's brief is clear: build agents that move beyond the chatbot to reason, plan, and act under your oversight. So we built Off-Duty for the people who need it most, an agent that runs a brick-and-mortar store while the owner is off duty, and proves every action in MongoDB.

What it does

Off-Duty doesn't just answer questions. It does the work, and it keeps the owner in control.

At the counter, customers just talk. Off-Duty takes the order grounded in live inventory, so it never sells what's sold out and never breaks the owner's pricing rules.
On the owner's console, it runs the store and shows its thinking. Ask it anything and watch the live agent trace route to the right specialist, then read straight from MongoDB. Every answer carries a grounding receipt you can click into.
Review-to-Action is the multi-step mission: one button reads a customer review, finds the product with Atlas Vector Search, checks live stock, drafts a reply, and queues the fix.
Shelf vision: snap a photo and Gemini reads the stock for you.
Human in the loop: nothing happens behind your back. Every action waits in "Needs you" for one tap to approve, reject, or undo.
Evidence drawer: trace any action back to the exact MongoDB documents it touched, by trace_id.

How we built it

A Google ADK multi-agent system: one supervisor that reads intent and delegates to ordering, inventory, vision, and owner specialists, built with Google Cloud Agent Builder. The brain is Gemini 3 on Vertex AI (with gemini-embedding-001 for 768-dim vectors). It serves from FastAPI on Cloud Run, streams the agent trace over SSE, and runs entirely on MongoDB Atlas: aggregation with query plans, Atlas Search, Vector Search, and an RRF hybrid. The MongoDB MCP server is the partner superpower, running live and read-only inside the app, with every call written to the evidence trail. Every read, write, and recommendation is auto-logged to agent_action_logs with a trace_id, and we reviewed the whole thing against the Google Cloud Well-Architected Framework's Security pillar. Review data was seeded from public datasets (Hugging Face yelp_review_full and a Kaggle restaurant-reviews set); the product uses Gemini for multimodal reading and trains on nothing.

Challenges we ran into

Making delegation real instead of one mega-prompt. Running MongoDB's MCP server live inside a request path. And the hardest one: building a trust layer where every answer is grounded and every write is auditable and reversible.

Accomplishments that we're proud of

A deployed agent that acts, summarizes, and verifies, with a live evidence trail, hybrid retrieval, human-in-the-loop control, and a security posture aligned to Google's own framework.

What we learned

Grounding plus an auditable trace is what turns an LLM demo into something an owner will actually trust. And the fastest way to make a slow agent feel fast is to stream its real thinking instead of hiding it behind a spinner.