← Back to the project post on chuckeesley.com

Proxy-Expert Practice EARLY-STAGE DEMO

A simulation of the three-agent system from the Stanford Accelerator for Learning project. The expert below, "R. Haddad," is a fictional composite written for this demo; no real DCI fellow is represented.
Early-stage ideas. The whole project, including this demo, is still at the ideation stage, and the design will change as we learn. We are starting with a first round of testing and training with a select group of DCI fellows, followed by a pilot in E145 in Winter 2027. Feedback and ideas on any part of it are very welcome. Get in touch.
This copy of the demo shows the stored materials: the expert's problem representation, the discovery decision log, the verification probe bank, and a worked example of the listen-in coach. The live AI responses (running the Discovery Agent, interviewing the proxy, generating fresh probes, and coaching a pasted transcript) are in the live version on Claude, which needs a Claude account.

The problem we saw in students

Students can read The Mom Test and work through customer discovery, then come back from a dozen interviews with a polished slide that sounds right and isn't. They rarely get to practice against someone who knows the domain well enough to refuse a shallow framing, and when they do, the social gap keeps them from asking the naive question or being wrong out loud. We saw the same pattern with industry innovation teams in the Saudi and ITRI programs, where the gap was between a technology idea and an operational problem someone senior had lived with for decades.

1 · Discovery Agent
Runs multi-session interviews with a real expert. Decides when to probe for a specific instance, when to flag a thin region, and when to schedule a re-interview to reconcile contradictions. Output is the expert's problem representation, which the expert reviews and can veto.
→
2 · Proxy-Expert Agent
Embodies that representation. Students interview it as if it were the expert. It tolerates repeated failure, answers good questions with concrete stories, pushes back on pitches and hypotheticals, and keeps the expert's own uncertainty intact.
→
3 · Verification Agent
Generates plausible-but-wrong restatements the student must tell apart from accurate ones, so nobody passes by echoing vocabulary. Difficulty adapts to demonstrated depth.
→
4 · Listen-in Coach
During the real interview, the proxy listens in and suggests good follow-up questions, what to probe, and where the team's understanding of the problem is still weak or inconsistent with what the expert has said.
→
Transfer
Blind raters score question quality and problem comprehension in the real interviews. Transfer to the live conversation is the outcome we care about.

How the study uses it

ArmWhat teams get
Full systemProxy-Expert practice plus adversarial verification, alongside the course's normal customer-discovery sequence.
Ablated (candidates, pre-registered)Verification removed; or proxy interaction reduced to reading the problem brief; or lower-fidelity proxy built from less interview depth.

Random assignment at the team level across E145, MS&E 272, and the international professional-development program. Every team gets a working tool, only the version differs.

Where it generalizes

Any setting where the scarce resource is access to someone who knows the domain: corporate and government innovation teams, engineering capstones, design studios, clinical and legal training. The demo uses an industrial maintenance expert because that problem space came up repeatedly with the international industry cohorts.

Expert: R. Haddad fictional composite

30 years in plant maintenance and reliability; last 9 as VP of Maintenance at a mid-sized specialty-chemicals company with three plants. Now in a mid-career transition fellowship.
3 discovery sessions · 2h 40m total · representation reviewed and approved by expert

Problem representation

Agent decision log (from the sessions)

Try the Discovery Agent

Paste an expert's answer and the agent decides its next move. Try a vague answer, then a concrete one.

Interview the proxy expert

You are a student team exploring a possible venture in industrial maintenance. R. Haddad has agreed to talk with you.
Try a weak question first to see the pushback:

Practice coach

Scores each of your questions against interview best practice. In the study this feedback is part of the full-system arm.
Session tally
No questions yet.

Is this restatement of Haddad's view accurate?

Some are right. Some use the right vocabulary and get the substance wrong.

Demonstrated depth

Current difficulty: 1 of 3
Two correct in a row moves you up a level; a miss moves you down. Level 3 probes distort cause and effect, overstate the expert's certainty, or credit a fix the expert already tried and abandoned.
Record
0 of 0 correct

The proxy listens in on a real interview

Once a team sits down with the real expert, the proxy can follow along. It compares what the team is asking and concluding with the expert's approved problem representation, and quietly suggests good follow-up questions, things worth probing, and places where the team's understanding is still weak or inconsistent. It never speaks in the interview. Below is a worked example from a team's interview with R. Haddad (fictional).
Follow-up to askProbe hereWeak or inconsistentGood move
Team's working problem statement:

Try the listen-in coach

Paste an interview transcript so far (one turn per line, starting with "Student:" or "Expert:") and the team's current problem statement. The coach suggests what to ask next.