← AI Work In Production

Case Study 01 — Legislative Intelligence

Kentucky Legislative Session Monitor

Automated monitoring of the 2026 Kentucky General Assembly — ingesting every committee meeting, surfacing enterprise story opportunities, and delivering editorial intelligence directly to reporters at scale.

131
Committee meetings ingested
469
Story ideas surfaced
38
Stories that aired
$1.65
Cost per aired story

The Kentucky General Assembly's 2026 session ran from January 6 through April 15 — a 10-week sprint of committee hearings, floor votes, and policy debates across dozens of standing committees. On a typical session day, the General Assembly runs six to ten committee meetings simultaneously. Our newsroom could staff two of them.

The result is predictable: stories get missed. Legislative action on education funding, healthcare access, and public safety advances with little to no coverage, because the newsroom was late or never knew it happened.

A fully automated pipeline built in MindStudio monitored the Kentucky Legislature's YouTube RSS feed continuously throughout the session. When a new committee meeting recording is posted, the system ingests it without human intervention — extracting audio, transcribing it via AssemblyAI, and analyzing the transcript with a large language model prompted to think like a beat reporter.

The output is a structured editorial report delivered by email to the Executive Producer. Story ideas are reviewed, filtered for newsworthiness and broadcast relevance, and forwarded to reporters for assignment — preserving human editorial judgment at every step.

"The AI surfaces. The journalists decide. That division of labor is intentional — and it's reflected in the numbers."

MetricResultNotes
Committee meetings ingested131Full automated coverage
Story ideas generated469$0.13 per idea
Stories pitched to reporters46After editorial review
Stories assigned3883% pitch acceptance rate
Stories aired38100% completion rate
Total system cost$62.69$1.65 per aired story
Challenge 01
Hallucination Guardrails — Lawmaker Attribution
The Kentucky General Assembly does not announce legislators by name during committee hearings — members are addressed by district number. In early testing, the LLM inserted lawmaker names by cross-referencing training data to fill the gap. The solution was explicit prompt engineering: the system was instructed never to attribute a statement to a named individual unless that name was spoken verbatim in the transcript.
Challenge 02
Pipeline Architecture — Live Video Filtering & Duplicate Prevention
The RSS feed surfaces the five most recent committee meeting videos — but at any given moment, one or two may be live streams still in progress. A single-step filtering approach failed consistently. The solution was two discrete steps: first, identify and discard live streams; second, check remaining videos against previously processed IDs and route the oldest unprocessed video for transcription.
Challenge 03
Reporter Feedback Loop — Output Design
The system was built with reporters, not just for them. An early version led with story ideas followed by a meeting summary. Reporters found themselves scanning past the ideas to understand context before evaluating them. Inverting the format — summary first, story ideas second — was a small change with meaningful impact on usability.