About the project

Inspiration

Real-estate sales teams rarely fail because another dashboard is missing. They fail when a changed price list, payment plan, availability sheet, or policy reaches people at different times—and nobody can prove which answer came from which approved source.

Wooho began with a stricter question: what would an agentic system look like if AI could move fast but could never silently become the source of truth? The result is a source-governed sales-readiness fleet in which every proposal keeps its evidence, every high-risk change waits for human review, and every published answer can be traced back to an immutable source version.

What it does

Wooho implements a quarantined project-document ingestion boundary, anchored extraction contracts, and eight typed agent stages: intake, truth audit, curator, compliance, readiness, training, sales intelligence, and delivery.

The fleet can identify conflicts, prepare structured knowledge, propose training and field guidance, and ground Ask Wooho answers with claim-level citations. It cannot publish canonical truth. An authorized reviewer sees the evidence and version diff, then an idempotent, optimistic-concurrency command atomically promotes the exact reviewed version. The native Android client is built to consume only tenant-scoped, client-safe projections; public Play distribution and end-to-end push delivery remain open gates.

This hackathon build also proves retrieval as a real Google Cloud path: a private Cloud Run ADK runtime uses Gemini 3.5 Flash Lite through Vertex AI and a deployed Vertex AI Vector Search endpoint. A bounded bilingual validation run returned 100% recall@8 across 20 queries, rejected all seven stale/unsafe filter cases, leaked no tenant or project data, and cleaned its synthetic data afterward.

How we built it

Wooho is a monorepo with three deliberate surfaces: a public Next.js marketing site, a Next.js administrative workspace, and a native Kotlin/Jetpack Compose field application. A trusted TypeScript /api/v1 command/query boundary owns canonical writes. Firestore in eur3 stores approved truth and metadata; Cloud Storage isolates source objects; Cloud Tasks and Pub/Sub carry typed references rather than raw documents.

The Python runtime uses Google ADK with Pydantic contracts on private Cloud Run in europe-west4. Model identifiers come from an allowlisted Secret Manager registry. Gemini 3.5 Flash Lite performs generation and Gemini Embedding 001 supports retrieval. Every vector carries tenant, project, source/version, approval, language, client-safety, validity, and freshness metadata, and Firestore is rechecked after retrieval.

Infrastructure is declared with Terraform. A validation Android registration slice has provider proof for Firebase Auth and App Check; Cloud Monitoring resources capture runtime state, and append-only audit records preserve actor, request, version, citation, cost, and decision evidence. This does not claim full production enforcement across every client.

The four-minute submission film uses nine receipt-backed Higgsfield Text to Speech V2 generations with the MiniMax variant and the fictional Arthur preset. The repository preserves every provider result ID, source hash, derived timing-safe stem, cost, and final-media QC result. Product screenshots in the film and public resource page are real API 36 emulator/browser captures; Higgsfield supporting imagery is never presented as product evidence.

Challenges we ran into

The hardest problem was not calling a model; it was keeping retryable asynchronous work correct. An agent attempt can time out after a provider accepted it, a task can be delivered more than once, and a review can become stale while a user is reading it. We introduced deterministic task identities, leased attempts, request digests, cost reservation/commit/release ledgers, optimistic versions, and replay-safe audit events.

Cloud regions also forced an explicit architecture decision: Firestore remains in the European multi-region eur3, compute and Vertex AI run in europe-west4, while Cloud Tasks uses its supported europe-west1 queue. The product makes these boundaries visible instead of relying on defaults.

Finally, we resisted the easiest demo shortcut: fake customer outcomes. Synthetic data is restricted to labeled fixtures and validation tenants; the public surface does not present invented logos, metrics, inventory, billing, or testimonials.

Accomplishments that we're proud of

  • Deployed two private, Ready Cloud Run validation services for the ADK agent runtime and ingestion runtime.
  • Deployed a tenant-filtered Vertex AI Vector Search endpoint and deletion-protected named Firestore validation database.
  • Proved 20 bilingual retrieval queries at 100% recall@8, zero tenant/project leakage, and seven out of seven canonical safety denials.
  • Implemented exactly eight typed agent stages with model/prompt versions, citations, confidence, usage, retry, and review contracts.
  • Built a 40-screen native Android field-agent foundation with 63 JVM tests and a source-reconciled 67/67 API 36 instrumentation run.
  • Produced a 240-second, fully decoded submission master with nine provenance-bound Higgsfield Arthur narration stems and real API 36 product captures.
  • Kept the product honest: AI proposals never publish truth, provider outages fail closed, and production does not fall back to mock data.

What we learned

Agentic architecture is a governance problem as much as a prompting problem. Retrieval metadata is part of authorization. Idempotency is a product feature. Cost reservations need the same rigor as billing. Arabic evaluation cannot be inferred from English quality. And a compelling enterprise AI demo gets stronger—not weaker—when it clearly distinguishes implemented code, live provider evidence, and the gates still ahead.

What's next for Wooho

Wooho is a production-minded validation build, not a public GA claim. The next milestones are authenticated end-to-end task delivery, human EN/AR acceptance, Play-distributed App Check and physical-device proof, full recovery drills, effective legal policies, and an explicitly approved live billing lifecycle.

Built With

  • clamav
  • cloud-monitoring
  • cloud-run
  • cloud-storage
  • cloud-tasks
  • firebase-app-check
  • firebase-auth
  • firebase-cloud-messaging
  • firestore
  • gemini-3.5-flash-lite
  • google-adk
  • jetpack-compose
  • kotlin
  • next.js
  • polar
  • pub/sub
  • pydantic
  • python
  • secret-manager
  • terraform
  • typescript
  • vercel
  • vertex-ai
  • vertex-ai-vector-search
  • zod
Share this project:

Updates