A voice-based psychological assessment platform, engineered for research-grade AI
How Metricoid built an end-to-end platform where participants answer structured prompts by voice, and a custom AI pipeline transcribes, extracts acoustic signal, and classifies relational patterns — architected from day one for academic validation, HIPAA-readiness, and full research reproducibility.
“"Extract the signal from day one — but let classification earn its accuracy through expert-validated calibration, never from uncalibrated assumptions."”
— Metricoid engineering principle“"A research platform is only as credible as its method. We built this one to capture everything from the first recording, classify reliably from transcripts today, and switch on voice-driven scoring only once expert-validated data makes it scientifically sound — with every output traceable to the exact model and rubric that produced it."”
A physician-researcher with a new framework — and no platform to test it
A US-based behavioral health founder — a board-certified physician — had developed a framework for understanding people’s relational patterns from how they narrate their own lives. The hypothesis was promising, but there was no way to test it at scale: it needed a rigorous, academic-grade instrument, not a consumer app.
The question was whether AI could classify these patterns from voice responses to carefully designed prompts — and, critically, whether those classifications would hold up against independent expert judgment. That meant building a genuine research platform: invite-only cohorts, structured assessment protocols, built-in expert validation, de-identified data exports, and reproducibility rigorous enough to support publication.
Metricoid was engaged to design and build Phase 1 — the academic MVP — on a fixed, not-to-exceed scope delivered across six milestones. This case study covers the platform as delivered through Milestone 3: the full participant experience and the working AI classification pipeline.
Measuring how someone speaks — not just what they say
The research rests on subtle signal: not only the content of a person's answers, but the way they tell their story — pace, hesitation, pitch movement, vocal tension. Capturing that signal is straightforward. Turning it into a valid classification is where most voice-AI projects quietly fail.
A raw acoustic measurement is meaningless on its own. When the system records a jitter value of 0.012 from a browser microphone, it has no way of knowing whether that indicates vocal stress or is simply normal for that speaker on that device. Feed those uncalibrated numbers into a classifier and you get confident output built on nothing — which, for an academic platform, is worse than useless: it destroys the credibility the whole study depends on.
So the engineering problem had three parts at once. Capture every acoustic feature from the very first recording, so no data is ever lost. Classify reliably today from transcripts, so the platform produces real results immediately. And architect the system so that expert-validated calibration can switch on voice-driven scoring later — without re-engineering the pipeline.
On top of that: deeply sensitive personal data demanding HIPAA-grade handling, and version traceability strict enough that any result can be reproduced months later for peer review.
The key engineering decision: separate acoustic feature extraction (done from day one, at effectively no cost) from acoustic feature integration (earned through clinician-validated calibration). Capturing the signal early builds the dataset the research needs; using it prematurely would make the AI less accurate, not more.
Five stages from spoken answer to structured profile
Everything heavy runs asynchronously. The participant records and leaves; the pipeline processes in the background and notifies them when their report is ready.
Voice responses captured
The participant answers 88 structured prompts by voice, one at a time, with record / pause / re-record and multi-session resume. Each clip uploads directly to encrypted storage.
Transcription & feature extraction
On submission, a background job transcribes every recording with word-level timestamps and, in parallel, extracts the 20 acoustic features — speech rate, pauses, pitch, and vocal stress.
Four classification passes
The transcripts run through four sequential LLM passes, each with a detailed clinical rubric: attachment patterns, personality structure, instinctual orientation, and relational-capacity scoring across 8 domains.
Risk flag & quality control
A non-diagnostic risk indicator is computed, and every classification carries a confidence score. Anything below threshold is automatically flagged for expert review — the pipeline never hides its own uncertainty.
Stored with full traceability
Results are written as a single record tagged with the exact model, rubric, and pipeline versions that produced them — so any assessment can be reproduced later for research review.
Classification from narrative markers across childhood and relationship prompts, returned with a confidence score and cited evidence from the transcript.
Personality structure, instinctual orientation, and 8 relational-capacity domain scores — each with its own confidence value.
Any confidence below threshold sets an automatic review flag, routing the assessment to a human expert.
model version · rubric version · pipeline version · timestamp — stored with every result.
An end-to-end platform — participant experience, AI pipeline, and the infrastructure under it
Participant voice assessment (web)
Delivered · M2Invite-only sign-up, intake forms, microphone setup and audio check, and the 88-prompt recording flow with record / pause / re-record, follow-up probes, and multi-session resume.
Secure async processing pipeline
Delivered · M3Queue-driven background processing: speech-to-text transcription with word-level timestamps, plus parallel extraction of 20 acoustic features per recording, with retry and fault isolation.
Four-pass AI classification service
Delivered · M3A standalone Python microservice running four rubric-driven LLM passes, producing classifications with confidence scores, a non-diagnostic risk flag, review flags, and full model/rubric/pipeline version tracking.
Admin & research portal
Delivered · M2–M3Study creation and management, prompt-set configuration, participant tracking, AI-result review with confidence scores, and de-identified research data exports.
Clinician validation portal
Designed · Milestone 4A blinded labeling workspace — audio playback, transcript review, and structured rubric entry — producing the independent expert labels that calibrate the voice layer and validate the AI.
HIPAA-ready infrastructure
DeliveredRole-based access control enforced at the API layer, AES-256 encryption at rest, TLS 1.2+ in transit, pseudonymized participant identifiers, and write-once audit logging.
Why accuracy is earned through validation, not assumed
The milestones are sequenced deliberately. Classification works from transcripts today. The voice features are captured and stored now, but they only become classification inputs once expert-labeled data gives each acoustic measurement a validated reference point. That calibration loop is the method — and it is what makes the output defensible as research.
Pipeline live
Classification from transcripts; all 20 acoustic features extracted and stored from every recording.
Delivered & demonstratedClinician validation portal
Experts label the same recordings blind, creating the paired dataset needed to calibrate the acoustic signal.
NextPilot & calibration
Clinician-labeled data calibrates acoustic thresholds and integrates voice features into an enhanced model; accuracy improvement is measured.
PlannedGo-live
Voice-enhanced classification, testing, stabilization, and deployment readiness.
PlannedThese are target baseline ranges for the Milestone 5 pilot — to be measured against clinician labels, not yet achieved. They are set deliberately honest: a rubric-driven model with no labeled data starts here, and is expected to improve as validated data accumulates. Setting realistic baselines is itself part of building research-grade AI.
Sensitive data, handled on a minimum-necessary basis
Pseudonymized by default
Participants are identified by codes, not names. Email is stored apart from assessment data, and research exports carry no direct identifiers.
Only transcripts leave for classification
The LLM receives pseudonymized text only — never names, emails, or audio. The minimum necessary to perform the classification task.
Audio stays encrypted & private
Recordings live in a private, encrypted store reachable only through time-limited, signed URLs — never public, never directly addressable.
A six-milestone build, delivered on a fixed, not-to-exceed scope
Foundation
Technical architecture, database schema, and UX / user-flow mapping across all roles — reviewed and approved before development began.
Platform
The web assessment platform: sign-up and intake, the voice recording experience, secure audio storage, and the admin portal.
AI pipeline
Transcription, acoustic feature extraction, and the four-pass classification engine — built, integrated, and demonstrated end-to-end.
Validation
The clinician labeling portal and validation dashboard — AI-vs-expert comparison that produces the labeled data for calibration.
Pilot & calibration
Pilot study with participants; calibrate acoustic thresholds, integrate voice features, and measure accuracy improvement.
Go-live
Voice-enhanced classification, testing and stabilization, and deployment readiness.
A working platform and a research-grade AI pipeline
Through Milestone 3, the platform delivers the complete participant journey and a functioning classification pipeline — built to the architecture approved at the outset, and demonstrated end-to-end.
What makes it research-grade is not a headline accuracy number — it is the discipline underneath: everything captured, nothing claimed before it is validated, every result reproducible.
Scope note: classification currently operates on transcripts. Acoustic features are captured and stored, with calibrated voice integration scheduled for Milestones 4–5. The platform is a research and validation instrument — it does not provide medical or psychological diagnosis.
Foundation, platform, and AI pipeline — built and shown end-to-end
Across four structured modules, with follow-up probes
Per recording, stored from day one for later calibration
Attachment, personality, instinct, and 8-domain relational capacity
What an off-the-shelf tool could never have done
This is not a problem a generic product solves. The classification rubrics encode a specific clinical framework; the pipeline has to be reproducible for publication; and the data is too sensitive to route through consumer tooling. It had to be built.
One-size scoring; no notion of a bespoke clinical framework
Rubric-driven passes encoding the researcher’s own framework
Opaque outputs you cannot reproduce or audit
Every result version-tagged and reproducible for peer review
Data sent through consumer infrastructure
HIPAA-ready, minimum-necessary, pseudonymized by design
The real asset is the validated dataset being built
Every assessment adds paired data — transcripts, acoustic features, AI output, and (from Milestone 4) expert labels. That accumulating, validated corpus is what no competitor can copy, and what makes each future model better than the last.
Built HIPAA-ready from the architecture up
The technical safeguards are built in from day one. Organizational HIPAA compliance — BAAs, training, formal risk assessment — remains the client’s responsibility as the covered entity, and the platform is architected so that activation requires no rework.
AES-256 at rest
Audio & database encrypted
TLS 1.2+ in transit
No unencrypted channels
Role-based access
Enforced at the API layer
Pseudonymization
Codes, not identities
Write-once audit logs
Tamper-evident trail
Minimum necessary
Transcripts-only to the LLM
BAA-ready
HIPAA-eligible AWS services
Client owns the IP
Code, data & models transfer
The Metricoid capabilities behind this build
Custom AI & LLM Engineering
Rubric-driven classification pipelines, confidence scoring, quality control, and version traceability — built for reproducibility, not black boxes.
Voice, NLP & Acoustic Analysis
In-browser voice capture, speech-to-text, and extraction of speech-rate, pause, pitch, and vocal-stress features for downstream modeling.
HIPAA-Ready Cloud Architecture
Encrypted, pseudonymized, audit-logged, role-scoped infrastructure on HIPAA-eligible AWS — designed for sensitive health data from day one.
Have a research idea that needs a real AI platform behind it?
We’ll tell you honestly what’s buildable, how to sequence it so the science stays sound, and what a fixed, not-to-exceed scope would look like. 15 minutes, no commitment.
