Every accepted conversation becomes ten typed dimensions of structured intelligence.

The conversation is the raw material. The ten dimensions are the structure the platform imposes on every accepted conversation. The set is vertical-stable in shape: every vertical gets ten dimensions, with two or three replacements depending on what the workforce actually knows. Retail gets one set. Consulting swaps customer-segment behavior for engagement-economics tells. Restaurants swap brand-perception for shift-level operational variance. The number stays at ten.

Ten is a calibrated number. Five collapses the dimensions into each other and loses the comparability across markets. Twenty fragments the signal across too many fields, each underfed by any single conversation. At ten, a fifteen-to-forty-five-minute interview produces enough density per dimension to support cross-conversation aggregation, and few enough fields that the schema remains queryable as a single object. The calibration is empirical. The set was tightened across two years of work before the first commercial deployment.

The ten dimensions above are the retail set. For consulting, customer-segment behavior is replaced with engagement-economics tells (rate realisation, scope creep, partner leverage). For restaurants, brand-perception map is replaced with shift-level operational variance (how the same store performs across the morning, lunch, and evening rushes). For pharmacy, regulatory and compliance risk takes a larger weight, and the AI-readiness scoring is calibrated for a regulated dispensing context.

The general pattern: every vertical keeps ten dimensions and replaces two or three relative to the retail set. The schema is portable across verticals. The methodology of replacement is documented per vertical and lives with the deployment.

Most AI-interview systems process every conversation and filter afterwards. The downside is that bad conversations pollute the aggregates the dashboard reports on. The platform runs the opposite design. A deterministic gate runs before any LLM call: aborted calls (under sixty seconds, fewer than six messages, fewer than four cleaned transcript lines), and low-signal calls (fewer than two hundred user characters, or an average turn under fifteen characters combined with a longest turn under forty characters), fail the gate.

When the gate fires, the quality score is set to zero, the reason is recorded, and the pipeline exits. No LLM tokens are consumed on a conversation that will not produce signal. The thresholds (sixty seconds, two hundred characters) are empirical. They were calibrated on the pilot corpus. They are tracked as parameters and are tunable per vertical.

A second gate runs after the LLM pipeline. If the intelligence-extraction call flags a richness of “thin,” or the quote-cleanup call flags every verbatim as low-value, or the quality audit flags data-usability as “low,” the conversation is tagged as an outlier and excluded from the executive view. The record is preserved for audit. It is not deleted.

The engineering depth lives on the docs site (linked below). At a methodology level, the pipeline does five things in a specific order, with specific atomicity properties.

Call 1, intelligence extraction. The transcript is converted into structured pain points, opportunities, threats, surprises, contradictions, a respondent profile, and zero to eight key verbatims. Every verbatim must be a literal substring of a real worker turn. A hallucination filter verifies this after the call. An invented quote is deleted, never rewritten.

Call 2, quote cleanup. Mechanical work at a cheaper model. The verbatim quote is the legal record and is never modified. A display-long version strips fillers without rewording. A display-short version extracts the single strongest brand-opinion clause from inside the display-long. The Quebec-French register stays intact.

Call 3, perception mapping. Run separately from intelligence extraction so the model does not collapse the two into one pass. Raw first, structure second. A cross-contamination check runs after this call: any brand-perception field that mentions a competitor or product the worker never named is cleared and flagged.

Call 4, quality audit. An independent agent grades the interview agent against six protocol KPIs on every conversation. The audit prompt references the same protocol specification that configures the interviewer. Both sides ground in the same document, so the audit cannot quietly drift away from the protocol it is measuring.

Call 5, translation. Analytical strings are translated French to English so the dashboard can serve anglophone users. The source-language verbatim is never translated. It stays in its original language as the provenance record.

Deterministic scoring. No LLM. A formula combines richness (extracted intelligence counts, with caps), depth (turn count, longest turn length, duration), and execution (the audit’s six KPI scores) into a zero-to-one-hundred composite. The formula is auditable. A client can request the exact computation that produced any score in their corpus.

One conversation runs one pipeline. A failure clears all partial state and the retry starts from scratch. There is no partial enrichment record in the database.

The pipeline runs four controls that catch the failure modes most pipelines ignore. Prompt drift checking. Every system prompt is registered by SHA-256 hash. A deployed hash that does not match the registered version fails the call closed. Hallucination filtering. Every key verbatim is verified as a literal substring of a real worker turn before persistence. Cross-contamination checking. Perception-mapping fields that introduce entities the worker never named are cleared and flagged. The enrichment audit trail. Every LLM call writes a row to an append-only audit table: pipeline version, prompt version, model identifier, token counts, duration, status. The table is never deleted.

The mechanisms behind these controls are documented in detail on Trust & Security.