Back to Research
Case study/GalenAI × zeldaLabs/Solutioning exercise · May 2026

The persona,
compounded.

How GalenAI, already running a full AI learning stack built around India’s CBME framework, partnered with zeldaLabs to explore a new frontier: persona-based onboarding and adaptive podcast learning. A focused, short-duration solutioning exercise, ready to be scaled up: two new pilot features, one shared substrate.

A 24-dimension persona vector visualised as a constellation, with two surfaces feeding signals into a shared model. PERSONA 24-dim VELA · ONBOARDING PROBE 16–20 signals INTERACTIVE PODCAST +8–14 signals per episode GALENAI APP STACK · ZELDALABS PERSONA SUBSTRATE
Fig. 1 The persona constellation. Each node is one of 24 persona dimensions held on the shared zeldaLabs substrate, sharpened by every GalenAI surface that writes to it, read by every application that calls into it.

01 · Executive summary

A persona that arrives before the first lesson.

GalenAI has already built a full AI-powered medical learning platform for Indian MBBS students and universities: a Learning Engine, an AI Content Generation Engine mapped to the NMC’s CBME framework, and a suite of Tutor Modes (Explain, Socratic, Case-Based, Quick Q&A, Build Your Own Profile), plus an Assessment Layer and Clinical Cases. The product is live and serving students. GalenAI owns the medical-domain expertise, the pedagogy, and every layer of the application stack. zeldaLabs brings the persona-intelligence research and the multi-tenant substrate that holds a shared learner vector.

In May 2026, the two teams ran a focused, short-duration solutioning exercise to explore what deeper personalisation could look like, layered on top of GalenAI’s existing platform. Two pilot features were built and demoed internally: Vela, a conversational onboarding probe that captures persona signals during onboarding; and an Interactive Podcast that pauses for checkpoint questions and adapts the rest of the episode in real time. Both write to a 24-dimension persona vector held on the zeldaLabs substrate. These are pilot additions to a platform that already ships, not the platform’s foundation.

24
Persona dimensions on shared substrate
A multi-tenant substrate that compounds across every application reading and writing to it.
10M+
Synthetic faculty pool, day one access
Via the Synthetic Society Connector. GalenAI’s faculty UX picks from the pool on demand.

This document is a single argument in three movements: the opportunity the two partners set out to explore; the architecture of the pilot (two new features layered onto GalenAI’s existing platform, using the zeldaLabs persona substrate); and the potential impact for students, for GalenAI as a product, for zeldaLabs as a persona engine, and for the medical institutions both partners ultimately serve.

02 · The opportunity we set out to address

A learner the system never quite sees.

Medical students in India carry one of the heaviest content loads in higher education, study in fragmented minutes, and learn in a language no platform speaks back to them. The result is a market full of content libraries, limited adaptive teaching, and a challenging student-to-teacher ratio.

CONTENT LOAD Thousands of pages of core MBBS text plus references, guidelines,high-yield PG synthesis on top. FRAGMENTED TIME 8AM 8 AM to 12 AM wards, lectures, study past midnight commute, OPD gaps, mess queue: tooshort for textbooks, ideal for audio. VERNACULAR GAP EN / हिं 65–70% prefer Hinglish to English-only English is the second language ofmedicine in India; no major platformspeaks back this way.
Fig. 2 Three structural constraints the existing market does not address.

Content delivery is not the same as teaching.

Traditional ed-tech platforms deliver content. They do this well. What none of them do is teach. A pre-recorded lecture on the Krebs cycle does not slow down when a particular student is struggling. The pacing does not change for a learner who responds better to clinical stories than to schematics. The system has no memory of how that learner answered yesterday, let alone six months ago.

That gap is not a feature gap. It is a data gap. To teach adaptively, a system needs a model of the learner that is persistent, multi-dimensional, and sharpens with every interaction. Building that model is the work. Once it exists, the teaching falls out of it.

Why this was an innovation-lab collaboration.

Three reasons made a research partnership the right shape for this work, with each side doing what it is positioned to do best:

  • Speed of exploration. A 24-dimension persona substrate and Hinglish-quality voice synthesis at production scale represent many quarters of platform research. With zeldaLabs as the research partner on that work, GalenAI got real learner signal back into its iteration loop in days, not quarters.
  • Compounding by design. A persona substrate lives where it can be shared. Holding the vector inside a multi-tenant platform means every application sharpens the same model. GalenAI’s medical-domain calibration today improves the substrate the next vertical partner will read from tomorrow, and the next partner’s signals improve it back.
  • Specialisation where it matters. GalenAI already runs the layer no platform vendor can build: their CBME-mapped Learning Engine, AI Content Generation Engine, Tutor Modes (Explain, Socratic, Case-Based, Quick Q&A, Build Your Own Profile), Assessment Layer, Clinical Cases, and Voice Layer. zeldaLabs runs the persona-modelling research that no single vertical would invest in alone. The solutioning exercise explored what happens when both stacks talk to each other.

03 · How we built it together

Where the two stacks meet.

GalenAI’s full application platform already ships: Learning Engine, CBME-mapped Content Engine, Tutor Modes, Assessment, Clinical Cases, and Voice Layer. For this solutioning exercise, two new pilot features (Vela onboarding and Interactive Podcast) were integrated with the four zeldaLabs persona substrate APIs, sitting on top of the existing stack.

GALENAI EXISTING PLATFORM PILOT FEATURES ZELDALABS PERSONA SUBSTRATE SHARED VECTOR GALENAI · 01 Learning Engine Pedagogy + adaptive lesson orchestration CBME-mapped, exam-track logic GALENAI · 02 AI Content Gen Topic-aware script & scenario production notes-to-episode, on-demand GALENAI · 03 Tutor Modes Explain · Socratic · Case-Based Quick Q&A · Build Your Profile Assessment · Clinical Cases · Voice existing stack · no Zelda integration SOLUTIONING EXERCISE · READY TO BE SCALED UP PILOT · 01 Vela Onboarding Conversational probe, 16–20 persona signals chat or voice PILOT · 02 Interactive Podcast Adaptive audio, checkpoint questions, Hinglish/English doc-upload → podcast Solutioning exercise integrates via persona substrate APIs ZL · 01 PersonaGen Engine 24-dim vector, archetype inference ZL · 02 Zelda Multimodal Signal extraction, per-dim deltas ZL · 03 ZeldaSpeak Multi-speaker voice, Hinglish at quality ZL · 04 Synthetic Society 10M+ personas, faculty + cohorts SHARED INTELLIGENCE SUBSTRATE · MULTI-TENANT 24 dimensions · 6 archetypes live reconciliation, cross-session continuity only the solutioning-exercise pilot features write to this substrate
Fig. 3 Four-layer architecture. GalenAI’s existing platform (top) is independent. It does not connect to Zelda. The two pilot features (Vela + Interactive Podcast) sit as a separate middle layer and are the only surfaces that integrate with the zeldaLabs persona substrate APIs below. All four substrate APIs write into the shared persona vector at the base.

Six profiles the model classifies against.

The substrate returns a blend rather than a single type, and the archetype drives the next product decision: which faculty to pair, which depth to open with, whether to lead with schema or story.

Anxious Planner over-organises; wants scaffolding Deep Mapper builds concept maps; follows tangents Schema-First structure before detail; frame, then fill Story-First learns via narrative & analogy Outcome Driven exam-anchored; high-yield first Social Learner peer dialogue; group dynamics
Fig. 4 The six archetypes from PersonaGen Engine. Names sourced verbatim from the Solutioning Brief, §3.1; one-line descriptors are editorial paraphrase.

What a single learner’s vector looks like.

Each learner is a point in 24-dimensional space. The shape sharpens with every interaction, and every application on the platform reads from this same shape.

COGNITIVE STYLE PACE & FOCUS MODALITY AFFECT & DEPTH SAMPLE PROFILE · REPRESENTATIVE, NOT MEASURED 24-dimension persona vector Schema-first NEET-PG aspirant after Vela + 1 podcast episode CURRENT VECTOR DIMENSION VERTEX
Fig. 5 A representative persona vector. Axis count (24) and quadrant groupings drawn from the brief. Individual axis names and the sample values are illustrative; the source does not enumerate them.

The four zeldaLabs APIs, put to work in the solutioning exercise.

API · 01
Persona model

PersonaGen Engine

The centre of gravity. Owns the 24-dimension vector, archetype inference, and reconciliation logic that keeps the vector consistent across sessions and surfaces.

  • Single learner state, persistent across applications
  • Confidence-weighted reconciliation prevents noisy over-correction
  • Live archetype classification: Anxious Planner, Deep Mapper, Schema-First, Story-First, Outcome Driven, Social Learner
  • Per-track weighting for NEET-PG and USMLE preparation
API · 02
Signal extraction

Zelda Multimodal

Touched most frequently per session. Accepts free-text, voice, or structured input and emits structured persona deltas with rationale traces.

  • All 24 dimensions scored in a single pass per reply
  • 13 specialised extractors layered for exam-track idiom
  • Reasoning trace returned with every delta, auditable from the application UI
  • Sub-second latency across text and voice inputs
API · 03
Voice synthesis

ZeldaSpeak

Multi-speaker voice rendering with per-persona acoustic profiles. Makes the podcast feel like a person teaching, not a TTS read-aloud.

  • Single or dual-faculty episodes with believable hand-offs
  • Hinglish + English at production quality; Indian-language roadmap
  • Acoustic features (pace, hesitation, uncertainty) returned to Multimodal
  • LINEAR16 output for direct streaming, no on-the-fly stitching
API · 04
Persona warehouse

Synthetic Society Connector

Queryable 10M+ synthetic-persona pool. Supplies the faculty roster on demand for every podcast episode GalenAI generates. The pool is a shared platform asset that GalenAI’s faculty-selection UX builds against.

  • Psychometrically diverse, demographically calibrated personas
  • Faculty sampling balanced across attributes the caller specifies
  • Per-session caching for re-pick without re-querying
  • Cohort generation for institutional callers (peer-study simulation)

The moment the learner picks their teacher.

The Synthetic Society Connector returns ten candidate faculty for the topic, balanced across attributes GalenAI’s application requests. The learner taps one; in dual-faculty mode, two. The most visually distinctive moment in the podcast UX, and the one most likely to be screenshot-shared.

SELECTED AI Dr. Ananya Iyer 38 · Clinical Psychologist Bengaluru warm structured code-mix Hinglish ► PREVIEW VOICE VM Dr. Vikram Menon 45 · Internal Medicine Mumbai authoritative concise pure English ► PREVIEW VOICE PS Dr. Priya Sharma 32 · Endocrinologist Delhi curious narrative Hindi-leaning ► PREVIEW VOICE
Fig. 6 Faculty selection. What the learner sees after the Synthetic Society Connector returns candidates. Card schema (name, age, occupation, location, three attributes) sourced from Solutioning Brief §5.1. “Ananya Iyer” appears in the existing demo; other faculty are illustrative synthetic personas.

Under a second, end to end.

Every learner action runs the same five-step chain. The whole loop is gated on the learner’s next move, so latency is invisible to them.

STEP 01 Learner replies text, voice, or tap STEP 02 FastAPI gateway routes to Multimodal STEP 03 Multimodal extracts deltas + rationale STEP 04 PersonaGen merges vector + archetype STEP 05 Radar updates + next turn queued END-TO-END LATENCY < 1 SECOND · GATED ON LEARNER’S NEXT ACTION
Fig. 7 The signal chain. Repeats once per learner turn on both surfaces.

Two surfaces, one persona contract.

SurfaceWhat the learner doeszeldaLabs APIs in use
Vela · Onboarding Probe6–8 short turns in chat or voice. Captures initial signals across 24 dimensions.PersonaGen Engine, Zelda Multimodal (+ ZeldaSpeak in voice mode)
Interactive Podcast4-min episode in Hinglish or English, multi-speaker faculty, checkpoint questions every 2–3 min that adapt the rest of the episode.All four: PersonaGen, Multimodal, ZeldaSpeak, Synthetic Society Connector
The cost shape, in a sentence. As GalenAI adds new Tutor Modes on top (tutor chat, viva sim, deeper clinical-skills modules), the persona substrate cost grows in step with usage, not with headcount or infrastructure. Linear costs, compounding intelligence: the unit economics improve as GalenAI’s product surface widens.

Concept to demo-ready in a short duration.

What made the build tractable wasn’t scope reduction. It was the partner’s ability to translate product vision into an executable technical contract in a single sync.

DAY 0 First sync zeldaLabs × GalenAI team SAME SYNC Scoped brief voice, persona, adaptive logic identified BUILD PHASE Voice · persona · interactive layer Hinglish pass, 10 faculty archetypes, checkpoints, adaptive speech SHORT DURATION DEMO READY Vela + Podcast both wired to shared persona vector Build timeline Sync → scoped brief → multi-surface solutioning exercise, in a short duration.
Fig. 8 The build. Calendar days from kick-off to a solutioning exercise that’s ready to scale.

Requirement-analysis speed

Vision to scoped contract in the first sync. Voice model approach, persona framework, quiz architecture, adaptive speed mechanism: all identified in one meeting.

Voice quality

Indian-accent TTS is hard; most models train on US/UK English. Hinglish delivery passed internal review as “feeling like a senior explaining it” rather than reading a script.

Persona engine, not personas

Faculty are generated on demand from content plus archetype plus style descriptor, rather than picked from a fixed library. Scales to any faculty type as the platform grows.

04 · Business impact

Four ledgers, one shared upside.

The same architecture creates value for four different stakeholders: students, GalenAI, zeldaLabs, and the medical institutions both partners ultimately serve. Each side reads from the same shared persona substrate, and each side sharpens it back.

GTM motions
2 channels
D2C learner subscriptions and B2B institutional licensing, both running on the same product stack.
Revenue compounding
Multi-year LTV
One signup, four exam cycles. Vector continuity removes re-onboarding friction at every transition.
Engineering leverage
Two stacks
GalenAI runs the application engines (Learning, Content Gen, Tutor Modes). zeldaLabs runs the persona substrate. Each side ships against the layer it knows best.
Defensibility
Grows on read
Every signal sharpens the same vector. Switching cost rises with every module that joins the fabric.

For students · engagement that compounds into retention.

Projected from internal review sessions, comparable studies in adaptive audio learning (Roediger & Karpicke, 2006), and analogies from engagement-heavy education platforms. A structured cohort test with 50–100 MBBS/PG students will validate.

BASELINE (CURRENT) TARGET POST-PILOT Daily active learning avg. minutes per day 8–12 min 22–28 min Session completion % of episodes finished end-to-end ~45% 70–75% 48-hour retention recall quiz score, same topics ~42% 58–65% 24h re-open rate “leave on a question” mechanic ~28% 45–52% Comprehension ease self-reported, 1–5 3.1 4.1–4.3 PROJECTIONS · PILOT N=50 (MBBS) + N=100 (PG) · 4-WEEK STRUCTURED TEST
Fig. 9 Projected pilot KPIs vs current baselines. Validation cohort planned.

For GalenAI · deeper personalisation on top of an already-live platform.

GalenAI already ships a full AI learning platform: Learning Engine, CBME-mapped Content Generation, multiple Tutor Modes, Assessment Layer, Clinical Cases, and Voice Layer. The persona-substrate solutioning exercise explores what becomes possible when those existing surfaces begin writing to and reading from a shared learner vector, opening the door to personalisation that starts on a student’s first second on the platform and sharpens across every session.

  • Time to learner benefit. A working teaching surface in days, with platform-level persona research happening on the substrate at the same time.
  • Defensibility surface. GalenAI’s medical-domain extractors, exam-track calibration, content generation, and faculty design are the layer that benefits most from medical-edtech specialisation. That is where GalenAI’s engineering compounds.
  • Capex shape. The persona substrate scales with usage, in step with learner adoption. Platform costs grow proportionally with the surface area GalenAI ships.
  • Continuity, by design. A returning Step 1 candidate’s persona vector from six months ago is already on the substrate. GalenAI’s episode opens calibrated on the first second. Continuity is something both partners’ engineering enables together.

For zeldaLabs · a working proof of the platform thesis.

GalenAI is, for zeldaLabs, the first real-world demonstration that a serious vertical application stack (with its own Learning Engine, Content Gen, and Tutor Modes) integrates cleanly with the zeldaLabs persona substrate. The application keeps its application IP. The substrate keeps its persona IP. Both meet at the API contract.

All four APIs exercised in one application stack

GalenAI’s Vela mode exercises PersonaGen and Multimodal. The Interactive Podcast exercises all four substrate services. The substrate scales linearly as GalenAI adds new Tutor Modes on top.

Reference architecture for what comes next

The same contract (application engines on top, persona substrate underneath, shared vector at the core) transposes onto adjacent medical applications: viva preparation, board-prep platforms, OSCE-style assessment, faculty-coaching modules, and future partners with their own engines. Each new application reuses the substrate.

The substrate gets denser, not just bigger

Every GalenAI signal sharpens the same vectors the next partner application reads from. Platform value compounds with each application onboarded, not just each user.

Auditable from the application side

Reasoning traces are surfaced in GalenAI’s application UI by default. Regulated education contexts (medical councils, board prep, accreditation bodies) get the explainability the substrate produces anyway, without additional engineering on the application side.

The compounding edge, in both directions.

The persona substrate compounds with every interaction. GalenAI’s medical-tutor and clinical-skills sim both write to it; partners joining the platform later read from a substrate that has already been shaped by GalenAI’s medical-domain signals, and write back into one that GalenAI’s next learner will read from.

SHARED Persona 24 dimensions GALENAI · LIVE Medical tutor writes via Vela & Podcast GALENAI · LIVE Clinical Skills Sim persona-aware practice cases FUTURE PARTNER Cohort simulator via Synthetic Society FUTURE PARTNER Board-prep platform writes back to substrate writes signal reads persona writes signal reads persona
Fig. 10 The compounding edge. Two GalenAI products write to the persona substrate today (medical tutor + clinical-skills sim, persona-integrated). Future partner applications read from the same substrate on day one and write back into it.

For institutions · what persona-aware learning at cohort scale could look like.

If the pilot validates, a medical college rolling out the partnership across a full cohort could gain persona-aware personalisation from week one of the academic year, without building any persona infrastructure itself. GalenAI handles the application layer (Learning Engine, Content Generation, Tutor Modes). zeldaLabs handles the persona substrate. The institution focuses on its own curriculum, faculty practice, and student-success operations.

Two adjacent doors open from there. Faculty pull synthetic peer cohorts via the Synthetic Society Connector to surround any single student with a realistic study group for a topic. And every additional learning modality the institution adds, whether GalenAI’s own Clinical Skills Sim or future partner applications on the platform, reads from the same persona substrate from day one. No student is re-onboarded.

05 · Use cases

Two motions, one platform.

The same persona model unlocks two distinct go-to-market motions: a direct-to-learner motion for individual exam aspirants, and an institutional motion for medical colleges. Both run on the surfaces shipped today.

Use case 01 · D2C · Individual learner

From first signup to returning student, the same vector for the whole lifecycle.

Three minutes of Vela onboarding (a pilot feature built on the zeldaLabs substrate) becomes the start of a multi-year exam-prep relationship, with the learner recognised from day one and across every return.

01GalenAI · Problem statement

Medical students in India face a multi-year exam journey across NEET-PG and USMLE preparation, and the existing market has no memory of who they are. Traditional ed-tech delivers content but does not teach: existing platforms don’t know who the student is on day one, can’t remember her six months later, and cannot carry her context from undergraduate study into PG preparation. Every exam cycle is effectively a cold start. The cost is low engagement (8 to 12 minutes per day, well short of the 22 to 28 needed for habit), high drop-off (around 45 percent session completion), and LTV that caps at a single exam.

02zeldaLabs · Solution approach

GalenAI’s Vela onboarding probe, built on the zeldaLabs PersonaGen and Multimodal services, captures 16 to 20 signals across 24 persona dimensions in three minutes. GalenAI’s Interactive Podcast then renders the first lesson already calibrated: a schema-first opener for a schema-first learner, paired with a faculty member whose teaching style matches, with checkpoints that adjust the rest of the episode in real time. Six months later, when the same student returns to prepare for a different exam, the persona substrate still holds her vector. The new episode opens calibrated on the first second. Both partners read from and write to the same substrate, so continuity is a property of the platform that GalenAI builds against.

03Business impact
  • Learning-style fit. Projected target: 60 to 70 percent of students report that the system understands their learning style well enough to plan their session by the third interaction. A session is whatever the student needs that day, and the platform meets them there.
  • Return engagement. Projected 45 to 52 percent 24-hour re-open rate (vs. ~28% baseline), driven by the “leave on a question” episode-end mechanic.
  • Multi-year LTV. One signup, multi-exam lifecycle. The same vector serves the same student across MBBS, NEET-PG, and USMLE preparation without re-onboarding friction at any transition.
  • Freemium funnel anchor. The first three minutes of onboarding is where personalisation is perceived, before any paywall enters the picture. That is the conversion moment.
  • CAC compression. “Feels like a senior explaining it” (verbatim from internal student review) drives word-of-mouth in WhatsApp study circles, which is the dominant referral channel in Indian medical cohorts.
Use case 02 · B2B · What the institutional motion could look like

What persona-aware cohort learning could look like.

A forward-looking sketch. The same architecture that serves an individual learner today could activate cohort-scale personalisation for medical colleges and coaching institutes tomorrow. This use case is not yet validated and is presented as exploration, not commitment.

01GalenAI · Problem statement

Medical colleges and coaching institutes want to offer personalised exam-prep to their cohorts. Two things stand in the way. They cannot reasonably build a persona substrate themselves: the underlying research and infrastructure work is multi-year and outside their core. And licensing existing content libraries does not solve the problem either, because content is what they already have. What they lack is a teaching system that adapts to each student and remembers them between sessions. Every new modality they roll out (simulation, viva preparation, faculty-led tutorials) currently means a fresh onboarding round for every student.

02GalenAI × zeldaLabs · What becomes possible

Imagine a medical college rolling out the GalenAI partnership to a full PG cohort at the start of the academic year. Every student runs the GalenAI × Vela onboarding probe in week one. The college builds on top of a persona substrate that GalenAI maintains the application layer for and zeldaLabs maintains the platform layer for. Each student’s first GalenAI podcast or clinical-skills practice case is personalised from minute one. Faculty can pull synthetic peer cohorts through the Synthetic Society Connector to surround any single student with a realistic study group for a topic. When the college layers GalenAI’s Clinical Skills Sim, viva preparation, or future partner modules on top, each one inherits the same persona substrate without re-onboarding any student.

03Why this matters commercially
  • Sales efficiency at cohort scale. One college close activates a full cohort of learners simultaneously. The institutional motion compresses per-learner activation cost meaningfully versus pure D2C acquisition.
  • ACV that grows with the cohort. Seat-based pricing on annual renewal cycles aligned to the academic year produces predictable revenue, with natural expansion as cohorts grow year over year.
  • Faculty enablement as a second product layer. The same persona signals feed cohort-level dashboards on archetype distribution, topic mastery, and at-risk learners. Faculty go from observers to informed interveners with no additional instrumentation work.
  • Stickiness that compounds with every module added. Each new GalenAI mode or partner application that joins the persona substrate inherits the cohort’s accumulated vectors. The switching cost rises with each addition, not because of lock-in, but because the value rises.

06 · What’s next

From solutioning exercise to pilot to platform pattern.

Validate

Structured cohort test: 50 MBBS students across 2–3 colleges, 4 weeks of access. Pre/post retention quizzes on matched topics; in-app analytics; exit survey on perceived quality and language preference.

Extend

100 PG aspirants (NEET-PG / USMLE) in Phase 2. Focus on high-yield topic comprehension and exam-readiness perception. Persona-vector deltas tracked across the four weeks.

Productise

Faculty persona generation moves from 10 demo archetypes to on-demand from content + style descriptor. Hinglish accent quality extended beyond metro dialects. Interactive quiz layer on by default.

Compound

Integrate podcast analytics back into GalenAI’s competency tracking. Begin scoping the next GalenAI Tutor Modes: tutor chat, faculty-led group sessions, viva simulation, each one reusing the same persona substrate.

“Two innovation partners, one shared substrate. GalenAI’s medical-domain signals sharpen the substrate today, and every partner application that joins reads from a substrate that has already been shaped, and writes back into one the next learner will benefit from.”From the joint partnership thesis, May 2026
The persona, compounded.
GalenAI × zeldaLabs · Joint case study · May 2026
Owners · GalenAI & zeldaLabs (joint)

Stay in the loop

Sign up for Lab Notes by zeldaLabs

Short dispatches from the frontier of synthetic human intelligence. No spam. Unsubscribe anytime.