Pulse

Pulse takes a proven organizational diagnostic that sat locked inside a DOS program for more than two decades and replaces its questionnaire with a condensed voice interview. Instead of filling in a long form, a leader talks for about 40 minutes with an agent that asks 25 questions and probes for context when an answer needs it. The conversation is then scored two ways: by a language model reading what was said, and by the original method's deterministic engine.

It started as a weekend hackathon that my dear friend Mo Sersouri organized: first, my friend Stephan cracked the old code, and I designed what the new experience could feel like over a weekend. A lot has happened since, and it is now live at getyourpulse.io as an invite-only alpha. We ran the interviews ourselves, then tested with a small control group of external participants. The diagnoses held up.

The thesis, in one line: kill the questionnaire, drive insights, and perhaps remove the need for the consultant altogether. The original idea was to democratize access to this kind of assessment for any mid-size firm, without the cost of a heavy engagement. And most importantly:

A condensed voice interview, scored on top of a deterministic engine, can surface what people do not write down: what they avoid, how they hesitate, what they reveal without meaning to.

A questionnaire captures what people are willing to tick. A conversation captures how they think. So instead of porting the form to a screen, we replaced it.

We settled on 25 questions, each with its own follow-up probe, designed to draw out long-form, natural, instinctive answers. Together they cover the ground the original questionnaire covered, condensed into an interview of about 40 minutes. The agent asks, listens, and probes for more context only when an answer is thin. A shorter 8-question version (about 10 minutes) exists for two reasons: to make testing easy, and to power the free tier.

That raised three questions this experiment set out to answer:

1. Can a 40-minute interview run at scale, cost-effectively? The consultant's time was the bottleneck. With a voice agent, the question becomes unit economics.
2. Is 40 minutes the right target? It is already a big compression of the original process. But a voice agent changes the trade-off between depth with one person and breadth across a team.
3. Do voice, speech and emotion calibration, scored on top of a deterministic engine, surface insights that stay invisible on paper? This is the core thesis.

Pulse is a working product with three audiences, each with its own surface.

Behind the interview is a full back end, not a demo. A consultant or an organization gets its own workspace, set up through a short onboarding that forks by audience. In it they create engagements, build a roster of respondents with personal links or QR codes, and watch each person move from not started to in progress to scored. Individual reports land as they finish. Once enough people have spoken, a group report compares them: where the team agrees, where it diverges, a playbook for each divergence, and an agenda for the readout meeting. Reports export to PDF. Accounts are invite-only, with referral invites and usage caps. Underneath sit an asynchronous scoring pipeline that can safely re-run any interview, and an operator console for accounts, costs and failures.

One design rule shapes every report: each claim cites the respondent's own words, shown inline as quotes, so anyone reading can audit a finding in seconds.

The alpha is invite-only. Accounts are minted by hand, each with a small assessment cap and a few referral invites, and every new account is seeded with a fictional sample engagement so the product makes sense before anyone records a word.

Yes on paper, about 5 to 6 times cheaper than the managed route, and essentially ready to switch on. The surprise was where the cost sits. The AI scoring is nearly free. The voice minutes are not.

Scoring costs cents. Across 23 logged production scoring runs, one individual report averaged about 4,000 input and 2,150 output tokens, roughly $0.04 at list pricing. The deterministic second pass adds about $0.04. A group report for a whole team costs about $0.07.

Voice costs dollars. On a managed voice-agent platform, a 40-minute interview runs about $0.11 to $0.16 per minute all-in, or $4.40 to $6.40 per interview. Worse, the plans sell concurrency through monthly minute buckets. Consulting demand is the opposite shape: an idle month, then 2,000 minutes in one week when an engagement launches.

The language layer is built and working. Calibrating the voice itself is the next layer.

Pass one: score the language, not the claims

The rule is “score what the language demonstrates, not what the respondent claims to believe.” So the first pass reads far more than the content of an answer: the words chosen, contractions, hesitations, fillers and self-corrections, silences, “I” versus “we,” past tense on a growth story, passive voice around failure, and which topics get volunteered or avoided. Every finding has to quote the respondent's own words, so any conclusion can be checked in seconds.

This is where a form goes blind. In our test group, one founder said plainly that they owned every decision, then went silent twice when asked about a hard decision and a recent failure. The report scored decisiveness as claimed but contradicted, and flagged that nobody else could actually decide. On the questionnaire, that founder would simply have ticked “strongly agree.”

Reports also sort the evidence into what the respondent already named, what is a blind spot (signal present but never acknowledged), and what is not on the radar at all. The blind spots are usually the most useful part.

Pass two: check it against the original scoring

Then the same conversation is scored against the original method. A model maps what was said onto the original questionnaire items, and the recovered weighted-sum engine scores those answers deterministically. When the two passes agree, confidence goes up. When they diverge, the divergence itself is shown as a signal.

Did it work?

We ran the interviews ourselves first, with anyone willing to test on their own business, then with a small control group of external participants. The diagnoses held up: people recognized their organizations. Across our evaluation set, every report carries 7 to 8 cited evidence items, and after several rounds of prompt changes all 5 evaluation cases kept the same primary stage and culture.

One honest caveat. When we replayed the original program's own two reference cases through the recovered engine, the verdicts did not fully reproduce. The likely causes are descriptor questions we do not score yet, industry-specific weight files we have not decoded, and a profile-matching step rather than a simple sum. Until that is resolved, the deterministic panel is labelled directional in every report, and the mismatch is pinned as a known-failing test so any fix announces itself.

Next: calibrating the voice itself

Today, hesitation is captured as it shows up in the transcript: pauses, fillers, broken sentences. What we do not measure yet is the sound: tone, pitch, pace, energy. Owning the voice pipeline (Question 1) is what opens that up, because the audio and the timing of every turn pass through our own server. The plan, in order:

1. Timing first. Response latency, pause length and speech rate per turn. Cheap and model-free.
2. Calibrate to the person, not a universal scale. Use the warm-up questions as each respondent's baseline, then flag deviations on the sensitive probes: power, failure, accountability.
3. Attach how it was said to what was said. Every quoted citation gains its delivery: said confidently, or said after a long pause.
4. Voice as evidence, never as verdict. The deterministic engine stays the baseline. The voice layer raises or lowers confidence and flags contradictions. It never scores anyone on its own.
5. Prove it helps. Blind review: do people prefer the reports with the voice layer, and do the flagged contradictions hold up in follow-up?

Guardrails come first: explicit consent to analyse delivery, no raw audio kept by default, no inference of protected traits, and no “lie detector” framing. The goal is calibration, not judgment.

Ordered by what unblocks what.

1. Voice cutover. Choose the voice (the hard part, for me), run one end-to-end short interview on the self-hosted pipeline, switch short interviews, observe, then switch long ones. Unblocks a measured cost per interview and access to the audio.
2. Make the reports palatable. Today they are dense and clinical: accurate, well cited, and a lot to take in. The next version leads with a plain-language summary, explains each section in a sentence before the detail, and trades raw 0 to 10 numbers for intensity words like high, moderate and low, the way the original consultant reports did. Independent of everything else, and it matters most to the people who read them.
3. Re-derive the deterministic engine. Decode the original program's own weight files, score the descriptor questions we skip today, and handle industry variants, until the original reference cases reproduce. Unblocks a baseline worth calibrating against.
4. Calibrate the voice, once 1 and 3 land: timing signals, then per-person baselines, then delivery attached to citations, then a blind review with consultants (see Question 3).