Product case study · 0→1 AI product

BERT: from behavioural concept to cross-platform beta.

Designing and launching a contextual behavioural-insights product while balancing user value, explainability, privacy, AI reliability and the constraints of a three-person contributor team.

RoleCo-Founder, Product & Technology
TimelineMarch–August 2026
Team3 core contributors
StageiOS & Android beta

The problem

Traditional workplace assessments can compress people into fixed labels even though behaviour changes with pressure, trust, ambiguity and team context.

BERT explored whether an adaptive mobile experience could capture behavioural evidence and turn it into a useful, shareable Blueprint without presenting it as a permanent personality type.

Product hypothesis: contextual behavioural evidence could create a more useful workplace reflection than a fixed label—if the experience remained understandable, trustworthy and specific.

My ownership

I translated the evolving concept into user journeys, roadmap priorities, requirements, user stories, acceptance criteria and API behaviour. I coordinated the product, design, strategy and engineering threads while remaining the primary backend and cloud delivery owner.

  • Product framing, MVP scope and prioritisation
  • Requirements, workflows and system contracts
  • AI integration and deterministic fallback strategy
  • Analytics, privacy readiness and beta release
  • Feedback synthesis and next-release direction

From hypothesis to beta

Discover
Approximately 12 target-user interviews conducted by a six-member UI/UX team informed the concept and journey.
Define
Converted findings and product strategy into onboarding, assessment, scoring, report, sharing and feedback workflows.
Build
Delivered Flutter clients supported by FastAPI, PostgreSQL, Azure services and versioned server-owned logic.
Launch
Prepared TestFlight and Google Play beta access, reviewer authentication, privacy disclosures and deletion workflows.
Learn
Combined journey state, structured survey evidence and free-text feedback into a prioritised product response.

Product and AI design

Context over labels

The experience models behavioural tensions, variation and contradiction rather than assigning a permanent personality category.

Evidence before language

Deterministic scoring creates the structured interpretation before Azure OpenAI assists with narrative expression. The language layer cannot become the source of behavioural truth.

Reliable fallback

AI output is schema-validated. If it is unavailable or invalid, the user receives the complete deterministic report rather than a broken or unsupported result.

Server-owned behaviour

Selection, scoring and report workflows remain versioned on the backend so that mobile releases do not fragment product logic.

What the beta showed

The controlled beta reached 27 registered users. Twenty completed a Blueprint and 12 submitted structured feedback.

83%Journey rated easy or very easy
92%Blueprint rated moderately or very accurate
100%Rated at least moderately useful at work
67%Likely or very likely to share

Positive usability and perceived value did not remove the central product problem. Users still experienced unclear birth-data justification, OTP friction, assessment repetition and reports that could feel generic.

Evidence → judgement → roadmap

Now
Improve OTP states, explain birth-data purpose, support journey recovery and make generation status reliable.
Next
Make Blueprints more specific, contextual, explainable and concise; review assessment wording and repetition.
Later
Strengthen sharing and invited-reader workflows only after the individual value loop earns trust.
Explore
Scientific validation, team comparison and enterprise administration remain gated by evidence and safeguards.
Product lesson: shipping a stable AI experience did not prove differentiation. The next challenge is to earn trust, reduce friction and show why each insight is specific and useful—not to add more features.

Responsible boundary and limitations

BERT remains a reflection and team-conversation aid—not a hiring or performance-decision instrument. The evidence is directional: 12 iOS respondents, early-adopter and self-report bias, no controlled experiment, longitudinal outcome study or formal psychometric validation.

Those limitations change the decision language: encouraging signals justify another learning cycle, not claims of product-market fit or scientific validity.

Open the two-page case study ↗Explore supporting artefacts →