← all projects

Harvest Automation

private

Agentic auditor for time bookings. An agent runs on a schedule, flags misbookings that leak billable revenue, and surfaces them to PMs in a review UI.

2026-04 – Present · @ Device Insight GmbH

  • Python
  • FastAPI
  • Anthropic Claude
  • SQLAlchemy
  • Alembic
  • PostgreSQL
  • Terraform
  • Azure Container Apps

An internal tool that puts an LLM agent to work on a real ops problem: PMs overseeing 50+ engineers' time bookings can't manually scan every entry for the misbookings that quietly cost the company billable revenue.

What it does

An agent runs on a schedule, audits bookings against a set of use-case rules, and writes findings to a review queue. PMs work that queue through a web UI — approve or dismiss — rather than reading raw timesheets. Two use cases (internal-vs-project misbookings; empty notes) are shipped; further audit use cases — unjustified customer bookings, cross-customer anomalies — are progressing behind them.

What I built

  • The agentic audit pipeline end-to-end — scheduled runs, the Anthropic-backed agent that classifies each booking with a confidence signal, and the FastAPI + Jinja review UI PMs actually use, including per-finding AI-cost and PM-approved-hours reporting.
  • The Azure deployment — Terraform IaC for Azure Container Apps + Flex Server Postgres with Managed Identity, plus the CI and pre-commit hygiene gates the repo runs on.

Source is private. This entry exists for context.