Nine projects, zero client names.

Data tracking external visibility existed; no system turned it into reports top management would use. The need was not a fixed template but a closed-loop platform that produces any report requested in natural language from in-system data, with sources shown.

From product definition to target architecture: system boundary, a 35-field master data schema, an "orchestrator + single-purpose module agents + report state object" prompt architecture, source/confidence/audit rules. "No unsourced claim enters a report."

Every week, thousands of records from 7 channels were cleaned by hand and placed into a 40+ slide deck; consistency depended on the person.

I built the path from raw CSV to deck as a modular Python pipeline; every rule is tied to a decision record and the pipeline reads values from records.

The prompt was 97.8% right when it said "spam" but caught only 23% of irrelevant content; in one period 3,901 records leaked into reports.

Root cause: a tie-break sentence overriding the platform rule. I rewrote the prompt around a hard phrase gate and built a 137-record, 7-layer regression set.

One label, two separate decisions. A single accuracy percentage conflated them; whose error is it, will it recur — there was no answer.

A protocol that measures the two axes separately, automatically catches label↔reason contradictions and checks determinism on its own; every finding has an owner.

"Is the vendor's spam analysis good?" needed a numeric answer; 2,525 records, 5 channels.

Six weighted dimensions: technical noise, off-brand content, same-name noise, duplicates, precision, decision consistency. Vendor output was compared record by record.

Rules lived in e-mail and memory; the AI read 82,000 characters for a single question and the same file held two different values.

A layered vault generated from decision records; router, lookup, a 22-rule linter, 12 benchmarks. Not a line was deleted.

Sending reports to specific people at specific times was manual; risk of forgetting and of wrong recipients.

Fixed-rule browser automation with no AI: job-folder model, a single recipient list, dry-run, evidence, logs.

Reviews were not matched to branches; "which branch is getting negative reviews" had no answer.

A matcher that generates query variants from the branch list and scores candidates on address/district/name tokens; results were joined to the sentiment table.

Reducing hundreds of brands' digital visibility to a single comparable score.

A data structure suited to scoring, score formula design, brand query sets, sector analysis tables.