Cru | Devpost

Inspiration

The informal economy in Mexico operates on trust, physical presence, and handwritten notes. Standard B2B SaaS platforms are structurally useless for people with low digital literacy. We recognized that to bridge this gap, the software interface had to be completely eliminated.

Cru was engineered specifically to empower female micro-business owners in Mexico, backed by Impeler, Apoyo Social, A.C. To maximize empathy, engagement, and adoption, the system operates entirely via voice notes in Mexican Spanish over Telegram. It provides these entrepreneurs with an autonomous back-office without requiring them to learn a single piece of software.

What it does

Cru is a multi-agent AI architecture that processes chaotic, informal audio dictation—including heavy slang, Spanglish, and background noise—and converts it into structured business operations. The architecture utilizes culturally resonant personas:

Doña Brújula (The Router): Handles intent classification and noise rejection, taking raw voice notes and outputting a strict JSON intent tag.
Doña Cuentas (The Ledger): Extracts structured accounting data from conversational dictation to log debts and partial payments.
El Lic (The Quote Generator): Converts messy verbal estimates into professional line items and invoices.
The Mentors (La Abogada de Hierro, La Maga de las Ventas, La Jefa de Piso): Provide specialized, jargon-free advice on finance, law, marketing, and logistics.

How we built it

The platform is built for high concurrency and low latency.

Core Engine: The backend is written in Go, utilizing an SQLite database (requiring CGO).
Infrastructure: It is designed to be deployed as a distroless container natively on Google Cloud Run.
AI Stack: We leverage Google's Gemini 1.5 Flash multimodal capabilities to directly ingest and parse the Telegram audio streams.
Telemetry: To strictly comply with XPRIZE tracking requirements, a specialized logger records all autonomous GenAI tool executions via a LogToolExecution() function.
Governance: The entire project operates under our Universal Code of Honor, enforcing human-centric alignment and radical transparency. Cru is Free Software, licensed under the GNU GPLv3.

Challenges we ran into

Building deterministic software on top of informal human language is highly volatile. The primary architectural challenge was intent classification and noise rejection.

When users dictate business operations, they ramble, mix intents, and use heavy Mexican slang. If a user says, "Anota que Don Chuy me debía 500 pero hoy me abonó 200", the system must distinguish between logging a new debt and updating an existing ledger state. Calibrating Doña Brújula to ignore the noise and output strict routing tags (like [LOG_DEBT] or [PRICE_CHECK]) without hallucinating required aggressive optimization.

Accomplishments that we're proud of

We are moving beyond the MVP phase to scale adoption within the Impeler network. By treating the AI not as a chatbot, but as an auditable, verifiable execution engine, we will expand the Go toolset to include automated inventory purchasing and deeper financial analytics, giving informal operators the same leverage as corporate enterprises.