Conecta · Talsa
People-and-culture platform for several companies of an agro-industrial group. Service tickets, staffing requests, forms, dashboards and a WhatsApp assistant.

- Role
- Main developer across backend, frontend, data pipelines and the assistant. A second developer joined in mid-2026
- Period
- Jan 2026 — Present
- Status
- In progress
- Stack
- Java
- Spring Boot
- Angular
- PostgreSQL
- Python
- FastAPI
- TimescaleDB
Conecta is the people-and-culture platform I build at Talsa, an agro-industrial company in Trujillo, Peru. It serves several companies of the same group and replaced calls, chat messages and spreadsheets with one system used by HR teams, area heads, field supervisors and workers at tablet kiosks. The code is private and the system holds personal data, so this page stays at the level of architecture.
What it covers
- Service tickets with a real-time Kanban queue.
- Staffing requests and disciplinary cases.
- Versioned dynamic forms and checklists that work offline.
- Uniform deliveries and a culture-voting campaign.
- Management dashboards over attendance, headcount and field data.
How it works

- Web system. An Angular app (standalone components, signals, zoneless change detection) on a Spring Boot modular monolith with PostgreSQL and Redis.
- Isolation between companies in two layers. Access is granted per company and module. On each request the API resolves the user’s access and writes it into PostgreSQL session variables. Row-level security policies read them and fail closed.
- Notifications. A transactional outbox, a worker that claims rows with
SKIP LOCKED, Redis pub/sub and WebSocket delivery. - Data platform. Python jobs pull HR and field data from vendor systems into a TimescaleDB warehouse and mirror curated tables into the app database. Loads are idempotent through content hashes. A window reconciler soft-deletes rows the source no longer returns, and aborts when a run would touch more than a set fraction of the window.
- WhatsApp assistant. A FastAPI service with a Celery worker answers workers and staff over WhatsApp. The language model chooses among a fixed set of tools. It never authenticates a user and never writes SQL, and each tool runs with the permissions of the person asking.
Engineering practice
- Database tests run on real PostgreSQL through Testcontainers, with contract tests for migrations and guard tests that enforce the backend layering.
- The assistant keeps no message text: sessions and short history live in Redis with a time limit, and the audit log stores a hash of the phone number.
- Releases follow semantic versions, with a deploy runbook that covers backup, rehearsal on a copy and rollback.