How CT Mirror Cut Meeting Coverage Time by 75 Percent with AI

A CT Mirror housing reporter needed to cover an eight-hour legislative committee meeting last week. Reading the AI-generated transcript took two hours. Listening to the full recording - with rewinding and scrubbing to find relevant moments - would have taken much longer.

The tool that made this possible: Legitalk, a new AI-powered system that transcribes Connecticut legislative hearings, indexes them by bill number and speaker, and lets reporters jump directly to the moments that matter.

Legitalk went live February 4, 2026 - just two weeks before the housing reporter's experience. It's an early validation of automated government meeting monitoring at scale. Last week, I talked via Zoom with Angela Eichhorst, CT Mirror's AI data reporter who built Legitalk. She walked me through how this tool works and how it fits into the AI strategy of the nonprofit news outlet that reports on Connecticut policy, politics and government.

Legitalk automatically pulls video from the YouTube channels of Connecticut's more than 30 state-wide legislative committees. Each video gets transcribed by Deepgram, then summarized by OpenAI's API with a specific focus: Every time a bill is mentioned, create a time stamp.

The result is a three-panel interface. On the left, reporters see bill-specific summaries. In the center, they can search the full transcript by keyword or speaker name. On the right, the video player lets them highlight any section of transcript and click to jump directly to that moment in the hearing.

"These hearings can be 10 hours long and have eight different bills discussed in different orders," Eichhorst explains. "The problem was journalists saying: I don't know where to find the five minutes that I care about."

The tool solves a recognition challenge that tripped up early versions. When legislators reference "Bill 1422," they might say "fourteen twenty-two," "one four two two," or "S.B. 1422." Eichhorst tuned the AI prompts with examples of all these variations so searches catch every mention.

Reporter feedback drove other refinements. The initial version let users highlight text and play just that clip. Reporters said no - they wanted continuous playback from that point forward. Eichhorst changed it within days.

Legitalk isn't built from scratch. Eichhorst adapted open-source code from Golden Dome (Vermont Legislature-focused code for SmartTranscripts from videos, based on Python and AWS Lambda) then customized it for Connecticut's specific committee structure and workflow needs.

"I think the real value-add is that I'm building tools specifically for the Connecticut Mirror," she says. "Not all news organizations work the same."

Legitalk isn't Eichhorst's only tool. Earlier, she built a system to triage Freedom of Information Act (FOIA) dumps - using AI to extract names from hundreds of PDFs and organize them into searchable case files. A reporter had received documents but didn't know how many people were involved or what the content covered. The common thread across Eichhorst's projects: using AI for volume problems that would take reporters days to handle manually.

That customization came from systematic user research. When Eichhorst joined CT Mirror in May 2025, she asked 15 colleagues a simple question: "Tell me what annoys you at work."

She got 85 pain points in response.

The biggest problems weren't technical - they were communication workflows and handoff processes between people. "A lot of people told me a tool they really liked in the newsroom was Airtable, which is just a form where we put in the budget of stories to manage content. A fix can be simple."

The sweet spot for AI, she found, was volume problems: "I have a lot of documents and I can't read them all" or "I don't know what's important in them" or "I can't be at a 10-hour meeting."

The housing reporter's experience wasn't Legitalk's first validation. In July 2025, before the tool existed, Eichhorst handled an urgent request that showed why it was needed.

A Connecticut representative was reading vulgar language from banned books during a legislative hearing - for shock value, as part of a book-banning debate. A reporter remembered this had happened before in an earlier meeting, but couldn't pinpoint when in a 12-hour video.

Eichhorst used AI to transcribe the video, then queried: can you find explicit language?

OpenAI initially blocked her. The content violated their policies. She rephrased and reprompted multiple times. Finally, she got results - verbatim quotes showing the representative had indeed done this months earlier.

The work took three to four hours. Listening to the full 12-hour meeting would have been impossible on deadline.

A critical detail: CT Mirror verified everything. They checked quotes against original documents and confirmed claims with secondary sources before publishing. The AI found the needle in the haystack. But humans verified that it was actually a needle.

This experience convinced Eichhorst that having searchable transcripts on demand would be valuable. "If that were to happen again this year, it would already be in my website, and the reporter could look for that representative's name and it would be in there automatically."

The scaling question looms. Connecticut has 169 municipalities that each hold multiple meetings. CT Mirror's 25-person newsroom can't possibly attend all of them. Eichhorst plans to extend Legitalk's approach to municipal meetings, but she's using the legislative tool as a prototype first.

The challenge will be signal-to-noise at scale. Legislative committees discuss eight bills over ten hours. Municipal meetings might cover zoning, budgets, school boards, and routine approvals with no clear indexing structure.

Eichhorst won't know the full feedback on Legitalk until Connecticut's legislative session ends May 5. "I still have to get full feedback from my colleagues on what they liked about the tool, what could be better," she says. The session is ramping up now with public comment hearings - the real test of whether the tool saves time when it matters most.

Can other newsrooms do this? Technically, yes. The Golden Dome code is open-source. Deepgram and OpenAI offer accessible APIs. And Eichhorst's user research methodology - asking colleagues what annoys them - costs nothing.

Practically, however, it's harder. CT Mirror has a two-person data desk and funding from the American Journalism Project that specifically supported Eichhorst's position. Most nonprofit and small for-profit newsrooms don't have that capacity. The economic reality of AI adoption in journalism: The tools that work best are custom-built for specific workflows, and customization requires dedicated technical staff.

CT Mirror isn't replacing reporters with AI tools. The newsroom is using AI to help their 25-person staff cover more ground. Eichhorst's work fits alongside the rest of her colleagues. She's built tools that save them time on tedious tasks so they can focus on actual journalism - talking to sources, crafting writing, spending more time on insights of data analysis.

Eichhorst is clear about what AI can't do. "AI can't report something new. AI pulls from the information that already exists. If there's an issue that no one has reported on, if there's an angle that no one has looked at, AI will not be able to pull that."

CT Mirror's AI ethics policy spells out the guardrails. The tool assists reporting. It doesn't replace verification. Humans check quotes, confirm claims, and make editorial judgments.

The housing reporter who got through an eight-hour meeting by reading the transcript still had to do the reporting. She still had to interview sources and write the story. Using Legitalk just meant she didn't have to spend more than 8 hours listening to find the 20 minutes that mattered.

Legitalk doesn't write stories. But it finds the moments worth writing about.