WhatsApp AI assistant
Operations staff get a chat assistant with multi-model failover, persistent memory, and knowledge-base search.
AI assistant + automation
Enterprise Tax Operations Platform
A WhatsApp AI assistant, a robotic document-acquisition pipeline, and a back-office operations dashboard for tax and customs workflows.
Ops platform
~120,000 lines Python · 1,664 tests
Overview
One platform ties together three jobs that used to be manual: an operations assistant on WhatsApp, a robotic pipeline that acquires documents from customs and tax portals, and a back-office dashboard for reviewing and exporting the results. The assistant routes between models with failover and keeps conversation memory and a searchable knowledge base; document handling is classical OCR, and the tax calculations are deterministic. AI-assisted templates are sandboxed and stay subject to human approval.
Outcome
~120,000
lines of Python
1,664
tests passing
41
versioned DB migrations
How it works
Robotic document acquisition
A headless browser pulls documents from customs and tax portals with per-account sessions and OTP / MFA relay.
Back-office dashboard
A Flask workspace covers multiple companies, job monitoring with live logs, and exports to Excel, e-invoice, and tax-return formats.
Outcome
A detail phase cut from 3m37s to 2m41s
Parallel fetching across four threads shortened the heavy detail step without touching correctness.
Built to survive failure
Multi-stage retries with exponential backoff keep long portal runs moving instead of stalling.
Runs on a real data layer
PostgreSQL 16 with connection pooling and a fail-soft dual-write between file store and database.
Want the same for your operation?
Tell me what runs by hand today. I will tell you whether it can be automated.