Green Peak Partners is a leadership consulting firm that coaches private equity executives, interviewing a leader's senior colleagues and provides feedback on the leadership team and how to grow to the next level. The sorting, theming, anonymizing, and drafting behind each report was all manual. We built the Insight Engine, a staged AI pipeline that ingests the transcripts from a new engagement and produces a finished report that surfaces deep insights that serve as valuable feedback to leadership teams.
In a typical engagement, Green Peak interviews senior colleagues about a leader, and then spends days reading transcripts, sorting hundreds of quotes into themes, scrubbing identifying details, and writing the executive summary and recommendations. The themes and resulting feedback are driven by direct quotes from real people, and the firm's credibility rests on the feedback being accurate and complete.
A quick AI shortcut doesn't work here. When Green Peak tested combining theme generation and quote extraction in a single LLM call, the model paraphrased quotes to fit the themes, and roughly 90 percent of them failed a word-for-word check against the transcripts. A paraphrased quote reads fluently, so nobody notices until a client asks where it came from. In a confidential document about a senior executive, one invented quote could sink the whole thing (and the relationship behind it).
What Green Peak wanted was simple to say and hard to build, upload the transcripts from a new engagement, generating insights using their expertise and proprietary body of work, with the quote sorting and theming already done, and without giving up the evidence standard the firm is known for.
The Insight Engine runs each engagement through a staged pipeline, goals first, then themes, verified quotes, anonymization, recommendations, and finally the executive summary. First inferring and then confirming the engagement goals is critical, otherwise the rest of the pipeline drifts toward whatever sounds interesting and loses sight of what the engagement was actually about. A few pieces of the build deserve a closer look.
The stages where a miss hurts most run Claude Opus and GPT-5 in parallel on the same prompt, and a third model call merges the two outputs, deduplicating equivalent findings while keeping anything unique either model caught. A theme missed early propagates through everything downstream, so the small extra cost of a second model buys real recall. There's even an admin diagnostic where an LLM judge checks whether the second model is still pulling its weight at each stage.
Every extracted quote gets verified by normalized substring matching against the source transcript, with a coverage score across its segments. Low coverage is rejected outright, and themes that end up with zero verified quotes get dropped entirely. This is also why themes and quotes are separate model calls, since combining them is what produced the 90 percent paraphrase failure. On top of that, the executive summary stage uses Anthropic's Citations capability, so generated claims come back tied to the exact source spans they were drawn from, and a grounding audit reruns the generation (up to three times) until every claim traces back to real evidence.
Before analyzing one person, the system scans every other subject's interviews in the engagement (chunked so nothing gets truncated) and pulls out verbatim passages that mention this person or describe the company as a whole, each attributed to its speaker. That widens the evidence base as reference-only context, never as a quote source, so attribution stays clean.
The summary and recommendation stages emit JSON that maps field for field onto Green Peak's Word template placeholders, so the pipeline renders a finished document with no reformatting work left for anyone afterward. A lenient parser plus retry handling keeps that contract solid even when a model gets sloppy.
Every change the consultant makes in the app (accepting or rejecting findings, rewording an anonymized quote, reassigning a speaker, editing recommendations, etc) is recorded with before-and-after values into a training stream. Those edits get distilled into rules and appended to future prompts, so the system leans toward how the consultant actually rewrites things. And the edits stick, because editing a stage forks a new run and only recomputes what comes after it, never the thing you just fixed.
A separate benchmark engine compares AI output to hand-curated golden reports using an Opus judge that checks whether each main point from the gold report is captured anywhere in the AI's version, crediting genuine coverage even when the AI groups findings differently and flagging new findings the human version missed. We keep the gold reports out of any generation prompt (with an automated check in the build process to make sure they never sneak in), so the score stays honest.
The Insight Engine is in production today, saving Green Peak consultants significant time and allowing them to focus on customer relationships and higher value tasks.
When we compared the initial output against the founder's hand-written report for the same engagement, it covered the same themes and recommendations, and we tuned it to generate additional coverage so they might catch things he might have missed. The wording and grouping differ, which is expected, but the important topics are all there.
The manual work the founder originally described, sorting hundreds of quotes and generating themes, now happens in the pipeline. His job shifts to reviewing and editing a strong draft instead of building one from scratch, and the stages assemble up to roughly 800 thousand characters of engagement material per call, reading no human wants to repeat for every project.
Since every edit the team makes feeds the learning engine, each engagement that goes through should come out a little closer to how Green Peak would have written it themselves.