[10]
AI PIPELINES — ASSECO EXECUTIVES
▓ TIER B · YEAR 2026 · STATUS: LIVE · LANGUAGES: PYTHON
Repeatable AI pipelines for managers — template-as-schema, zero hallucination
[FIG. 1] MISSION
For the AI Executives training at Asseco I built a set of pipelines a manager runs on their own project data with no IT rollout: daily-standup audio → meeting note, a Jira export → four-dimensional report, tender specs → risk map, a budget → FTE calculation. The engineering lives not in code but in prompt contracts: 14 szablon_*.md templates act as a rigid output schema — the agent neither adds nor omits sections — and the reporting standard mandates a Source column in every fact table and a (data)/(assumption) tag on every number, because budget decisions are made on these reports. A separate, deterministic branch is the local transcription pipeline: ffmpeg → Silero VAD → Whisper, with automatic backend selection per platform, so a meeting recording never leaves the machine. I packaged it all as four Claude Code skills plus PDF (ReportLab) and HTML (Chart.js) report generators with their own design system.
[FIG. 2] ARCHITECTURE
hover a block to see its description
[FIG. 3] CHALLENGES
[+][CH-01]
LLM-generated reports “drift”: different section order, different columns, fragments added or dropped — and with 14 distinct outputs produced by many training participants, that disqualifies the material. The fix turned out to be a contract, not a request: every prompt carries the same three-beat procedure — first read the template start to finish, then read the data, only then fill the template section by section, with identical headings, identical columns and metrics computed with the same formulas — plus a hard “do NOT add sections beyond the template, do NOT omit sections from it”. I iterated that clause into a form resistant to model creativity and repeat it verbatim in all five cases.
[+][CH-02]
A manager makes budget decisions in currency and headcount on these reports — a hallucinated number or an invented quote from a tender document is real business and legal risk. So I designed a standard where the report is auditable backwards: a mandatory Source column in every fact table, a (data)/(assumption) tag on every number, the formula with substituted values for every computed figure, and the rule “never state a cause as fact without evidence”, with a checklist validator verifying all of it before the report ships. In this standard, a hallucinated quote is explicitly defined as the gravest substantive error — not a typo.
[FIG. 4] AI LAYER
AI is the product here, not a garnish: Claude generates reports dictated by the “Common Rule” and the templates, and Whisper with Silero VAD transcribes audio locally. My job was taking freedom away from the model — template-as-schema forces identical structure across varying data, and provenance-first rules make every number in a decision report auditable back to its source.
[FIG. 5] GALLERY
SCREENSHOTS INCOMING — see SHOTS-WANTED