What it does
The driver photographs a receipt — diesel, tolls, per diem — and sends it to
a WhatsApp number. Likida reads it, validates the CFDI (Mexico's mandatory
e-invoice) against the tax authority, reconciles it against the advance and
the company's expense policy, flags what's missing, and returns a closed
settlement as a PDF and a line in the fleet's ERP. No app to install in the
cab, no training, no hardware.
Product principles
- The photo is not the source of truth; the CFDI is. The data model
keeps
comprobante_recibido (the photo — operational control, timestamp,
odometer) strictly separate from cfdi_validado (the XML verified
against the tax authority). No money decision is ever made on a photo
alone.
- A receipt doesn't get one verdict — it gets several. The same ticket
can be deductible for one tax and not another. The engine emits separate
verdicts, each one citing the specific legal basis behind it.
- The engine is deterministic. The LLM only extracts and drafts
language; every money decision, every deductibility call, every approval
gate lives in auditable, tested code. What can't be verified is marked
"needs confirmation" — the product never fabricates a number, a tax
regime, or a deduction.
Architecture
WhatsApp Cloud API → webhook → intake (vision OCR + CFDI validation) → reconciliation engine (deterministic) → settlement (PDF) → fleet's ERP
↘ specialized agents (finance, growth, leads, support, engineering…) ↗
| Module | Path | What it does |
|---|
| Intake | src/lib/likida/intake | WhatsApp photos → vision model → structured JSON + CFDI parsing and government validation |
| Reconciliation | src/lib/likida/cuadre | Reconciles expenses vs. advance + policy + tax rule → discrepancies, missing items, cited verdicts |
| Fiscal | src/lib/likida/fiscal.ts + normas/ | The tax rule engine — every counter cites its legal source and carries a test |
| Settlement | src/lib/likida/liquidacion | PDF generation + ERP export (CSV/JSON) |
| Invoicing | src/lib/likida/facturacion | Declarative portal automation (Playwright), queued via QStash |
| Agents | src/lib/likida/agentes | The agent runner: budgets, kill switches, dispatch |
| Conversational agents | src/lib/agents | WhatsApp conversation, prompts, mutation dedup |
| SaaS | src/lib/saas | Stripe, subscriptions, invoicing for Likida itself |
| Observability | src/lib/observability | Redacted logging, Sentry, fail-loud startup config checks |
Dashboards: /dashboard (the fleet), /admin (Likida's internal
console — cross-tenant, superadmin only), /chofer (the driver's portal)
and /demo (a live conversation simulator).
Likida doesn't replace the fleet's ERP — it feeds it.
Stack
| For | What |
|---|
| App | Next.js 16 (App Router) · React 19 · strict TypeScript |
| Data | Supabase (Postgres + row-level security + RPCs) — every schema change is a reviewed, tested migration |
| AI | OpenRouter — vision models for OCR, reasoning models for reconciliation, cross-provider fallback |
| Channel | WhatsApp Cloud API (Meta) |
| Documents | zxing-wasm (QR/barcode) · sharp · fast-xml-parser (CFDI) |
| Portal automation | Playwright + @sparticuz/chromium (cron-only) |
| Queues | QStash (signed enqueue + callback) · Postgres for the intake barrier |
| Output | pdf-lib · CSV/JSON export |
| Testing | Vitest (offline) + real-model harnesses in pruebas-manuales/*.prueba.ts |
| Hosting | Vercel |
Rules the code doesn't bend on
- Never fabricate a number. If real data doesn't exist, the screen says
what's missing and why. An estimate is shown declared, with its
assumption visible next to it.
- A label has to be true. If it says "this period," the query filters by
date. Demo data is marked as demo; test RFCs (Mexican tax IDs) are
fictitious but pass the real checksum.
- Fail closed, and say so. A database outage is never read as "no
settlements yet." A missing environment variable is screamed at startup,
not discovered live in front of a customer.
- Number formatting lives in exactly one place (
lib/formato.ts) — a
test fails the build if it shows up anywhere else.
Running locally
npm install
cp .env.example .env.local # fill in the keys — the file documents each one
npm run dev
Quality gates, in order:
npx tsc --noEmit -p . # 0 errors
npx eslint src/ # 0 errors
npx vitest run # everything green
npm run build # clean build
Without a WhatsApp number, the conversation simulator lives at /demo.
Synthetic demo data for local databases: scripts/demo/cinco-agentes/.
Deployment
- Vercel, aliased to
app.likida.ai. Builds are opt-in: the commit
subject must carry [deploy] (enforced via vercel.json's
ignoreCommand) — a plain push ships code without publishing it.
- WhatsApp webhook →
app.likida.ai/api/webhook/whatsapp.
- Scheduled jobs in
vercel.json; the invoicing cron requires
CRON_SECRET and queues through QStash.
- Runbooks:
docs/conocimiento/DEPLOY.md, docs/operacion/.
Documentation
docs/conocimiento/DEPLOY.md — production runbook.docs/conocimiento/CONFIGURAR-META.md — WhatsApp Cloud API, step by step.docs/operacion/ — runbooks of the agents, portals and operations.docs/legal/ — legal templates (DPA, SLA, security annex, retention).CLAUDE.md, CONTEXT.md, DESIGN.md — engineering rules, glossary and design system.
