Portfolio company data onboarding

From ambiguous source data
to reviewable financial metrics.

Invoice lines arrive in cents; headers use dollars. A CRM column called rev means potential bookings. This workflow checks the meaning, carries the evidence and requires review before publishing a financial result.

An AI-assisted portfolio prototype: synthetic financial workflows, typed CSV and read-only PostgreSQL snapshots, and local publication.

Watch the scripted replayRead the case study
Evidence before action

Unit mismatches and ambiguous revenue meanings become review items with traceable evidence.

Deterministic calculations

Generated monthly metrics execute in a sandbox and reconcile against reference calculations.

Separate reviewer gates

The agent cannot approve its own work. Certification binds the reviewed result to the published files.

Actual output. Replayable evidence.

Recorded stdout from the synthetic demo. Reviewer decisions are scripted; no backend or model API runs on this page. The replay metadata identifies its source revision, and the transcript contains the actual demo output.

Loading…
Loading the captured demo…
Read the full accessible transcript

A workflow you can inspect

Source profile → Mapping → Review → dbt + reconciliation → Certification → Publish

Database state survives process restarts. Approvals bind a content hash and expire when inputs change. Publication verifies the certified files and recovers across filesystem and database failure windows.

Explore the architecture and design decisions →

The audit that changed the checks

A ten-agent audit recorded 23 findings. One source used Revenue as an account category: a 100.00 credit became zero recognized revenue while every reconciliation check passed. Both calculators shared the same unchecked semantic assumption.

The repair makes canonical categories executable contracts. Unknown classifications block affected metrics; normalization requires a reviewed transform. Regression coverage also addresses rejected joins, revoked approvals and stale-worker publication.

My role

I chose the portfolio direction, reviewed recommendations and decided which capabilities to pursue. I used AI coding tools extensively for implementation, tests, audits and documentation. I present the prototype with its evidence and failures visible, and distinguish automated review from independent human validation.

Evaluation and boundaries

The SaaS demo publishes nine reconciled metrics. Typed CSV and read-only PostgreSQL extraction have synthetic end-to-end verification. A real historical retail exercise imported and profiled 10,000 rows; it did not certify financial metrics.

The original unfamiliar-schema probe produced zero proposals and matched 0/10 targets. Later frozen agent-authored corpora exposed additional safety failures. Repairs increased review coverage, with wrong proposals and substantial review effort still present. The failed results are preserved; remediation is feedback-informed, not untouched held-out validation.

Publication is local. There is no accepted live backend, warehouse deployment, independent human usability result or measured business savings. The public presentation runs on free GitHub Pages. Optional hosted deployment and the bounded local semantic executor have separate acceptance limits.

Continue through the project

Case study → Architecture → Evaluation and limitations → Release status

A historical 3:51 tour uses a clearly labeled synthetic voice and recorded fixture results. Six Claude Code sessions preserve actual tool calls and model responses; only one invoked a Skill body, so no general improvement is claimed.