Rodrigo Passarelli
kinn is a real-time Bayesian diagnostic interview engine for consultants, founders, and anyone running discovery calls.
Core trick
Scores candidate follow-up questions by expected information gain against a Bayesian belief about the stakeholder, updated after each answer.
Description
kinn is a real-time Bayesian diagnostic interview engine for consultants, founders, and anyone running discovery calls. After every stakeholder answer, Claude Opus 4.7 computes the Expected Information Gain of every candidate next question against a Bayesian belief over the stakeholder's actual situation — and asks the one that collapses the most uncertainty.
The frame is BED-LLM (Bayesian Experimental Design with LLM samplers): Opus 4.7 is *both* the answer-distribution sampler (predicting how a stakeholder might respond) and the belief updater (revising priors after each real answer). DSPy/GEPA compiles the prompts offline; runtime is a forced-tool-call loop with prompt caching, two-phase recompile, and dual-algedonic state separation.
Built solo across the hackathon week in three iterations (kinn → kinn2 → kinn3). The dual-gate benchmark publishes the honest result (0.852 vs 0.920 target) rather than ship a hidden failure.
README excerpt
# kinn — a Bayesian diagnostic interview engine  > **Built with Opus 4.7 hackathon submission · Apr 21–28, 2026** > Authored entirely within the hackathon window. Earliest commit on the kinn lineage: `b81880a` at **2026-04-22 10:16 +0100** (24 hours after the hackathon opened). The published repo's earliest commit is `32feb49` at **2026-04-25 01:06 +0100**. Full git history is the start-date proof. **Solo human author: Rodrigo Passarelli.** Claude Opus 4.7 used as AI assistant throughout (`Co-Authored-By` in commits — that's the whole point of this hackathon). > **Naming.** The internal Python package is `kinn3` because this is the third hackathon-week iteration: kinn → kinn2 → kinn3, all built within the Apr 21–28 window. The "3" is iteration depth, not "version 3 of pre-existing software." See [Iteration story](#iteration-story). **The problem.** Every consultant, founder, and product team starts client work the same way: a fuzzy 60-minute discovery call where they ask whatever comes to mind, write notes, and hope the gaps surface later. The expensive failure mode is asking the wrong next question — the one that locks in a wrong assumption. Real diagnostic interviewing is a skill that takes years; most people never get it. **What kinn is.** A real-time interviewing engine that, after every answer, computes the **E