Build Progress
A living record of design decisions, components, and progress. Updated as we build. Bookmark this page.
p0stman for ThynkNu / Last updated: 15 May 2026
01
Warm Guide, confirmed 23 April
Human, approachable, modern wellness. A friend who happens to understand insurance. Rounded corners, generous spacing, conversational tone. Think Headspace meets Monzo. Accessible without being childish.
Product
Kind + ant. Care, community, diligence. The platform users trust with their health insurance decisions.
Agent
Latin for “truth”. The conversational AI agent users interact with. Warm, human, bilingual-safe.
Display / 48px / Bold
Kindant
H1 / 36px / Bold
Your health, clearly understood.
H2 / 24px / Semibold
Insurance without the fine print.
Body / 16px / Regular
Vera is your personal health insurance guide. She understands your cover, remembers your history, and explains everything in plain language. No jargon, no forms, no call centres.
Caption / 12px / Medium
Rules engine verified. Last updated 23 April 2026.
Font: DM Sans (400, 500, 600, 700). Unified sans-serif with weight contrast for warmth. Modern, friendly, highly readable at all sizes.
02
Emerald primary, warm stone neutrals
Primary
#059669
Actions, agent accent, CTAs
Primary Dark
#065F46
Strong emphasis, depth
Background
#FAFAF9
Page background
Surface
#FFFFFF
Cards, panels
Agent Wash
#ECFDF5
Agent bubbles, highlights
Text
#1C1917
Primary copy
Muted
#78716C
Secondary text
Border
#E7E5E4
Dividers, card borders
Primary
#34D399
Actions, agent accent
Primary Dark
#059669
Strong emphasis
Background
#0C0A09
Page background
Surface
#1C1917
Cards, panels
Agent Wash
#052E16
Agent bubbles, highlights
Text
#F5F5F4
Primary copy
Muted
#A8A29E
Secondary text
Border
#292524
Dividers, card borders
Success
Warning
Error
Info
03
How Vera talks to users
Vera, just now
Simulated conversation showing Vera's tone, action cards, source labels, quick replies, and multimodal input.
04
19 components built, rendered live below
Policy Summary
Bupa International Premier, active until Dec 2027. Family plan covering 4 members.
HK$12,500
Annual premium estimate
Agent sm
Agent md
Agent lg
User sm
User md
User lg
I can see you have a family plan. Let me check what your outpatient coverage looks like for specialist visits.
Vera, just now
You, just now
Your plan covers specialist consultations at 80% after a HK$500 deductible per visit. Shall I find dermatologists in your network near Central?
Checking your policy terms...
Found 3 relevant clauses
Outpatient, specialist referral, and deductible sections
Could not verify provider network
The insurer's directory is temporarily unavailable
Tap any pill to edit. These show what Vera remembers about you.
Ready to submit your health assessment
Bupa-Premier-Policy-2026.pdf
PDF, 2.4 MB
I can see this is a Bupa International Premier policy, active until December 2027. Family plan covering 4 members with worldwide coverage.
This feature is part of the premium experience. Upgrade to unlock care navigation, claims assistance, and more.
This needs a human expert
This claim involves a dispute with the insurer that needs specialist review. I have prepared a summary of your case.
05
Three layers, zero guessing
Vera never answers from general AI knowledge. Every response is grounded in one of three verified sources. If none of them have the answer, she says so.
User
Asks insurance questions via text, voice, or document upload
Vera
Listens, calls the right source, explains clearly, cites everything
Rules Engine
DeterministicMichael's underwriting decision trees. Pure TypeScript. 234 conditions across all 15 categories. Every decision is testable and auditable. No AI involved.
Knowledge Base
VerifiedStructured domain knowledge: insurance glossary, HK market context, condition guides, policy terms. Vera reads and cites these files. Updated without retraining.
Refusal
Safety netIf neither the rules engine nor the knowledge base has the answer, Vera says so clearly and suggests next steps. She never guesses.
Why this matters: In health insurance, a wrong answer is worse than no answer. This architecture means Vera can explain a complex exclusion clause and trace it back to the exact rule that applied, or tell you she does not have enough information to answer. There is no middle ground where she guesses and sounds confident.
Vera is not a chatbot. She is an agent that reasons about your question, decides which tools to use, calls them, and synthesises the results into a clear answer. Here are the 9 tools she has access to today.
listCategories
Rules EngineLists all 15 disease categories and condition counts. Vera calls this when a user asks “What conditions can you assess?”
listConditions
Rules EngineLists all conditions in a specific category. When a user mentions “diabetes”, Vera searches the endocrine category to find the right decision tree.
lookupCondition
Rules EngineRetrieves the full underwriting decision tree for a specific condition: the questions that would be asked, possible outcomes (standard, exclude, decline, postpone), and the decision logic. Every answer is deterministic.
lookupKnowledge
Knowledge BaseReads verified knowledge files: insurance glossary (30+ terms), HK market overview, underwriting basics, condition guides, and what Kindant offers. Vera cites the source in her response.
evaluateCondition
Rules EngineRuns the deterministic underwriting evaluator from within chat. Takes a condition ID and user answers, walks the decision tree, returns a binding outcome (standard, exclude, decline, postpone). Auto-persists an audit log to Supabase.
readPolicy
Knowledge BaseReads user-uploaded policy PDFs that have been processed through the ingestion pipeline. Claude extracts structured data from the document. Vera can then answer questions about the user's specific cover, exclusions, and benefits.
escalateToHuman
EscalationWhen Vera encounters a question beyond her knowledge or a nuanced case that needs human judgement, she escalates to Sarah or Michael. An inline EscalationCard appears in the chat, and the case is routed to the admin dashboard via /api/escalate.
readDocument
VisionReads uploaded images (receipts, medical reports, insurance cards). Claude vision extracts text and key details automatically on upload.
Voice I/O
ElevenLabsBrowser speech recognition for voice input, ElevenLabs TTS for voice output. Vera speaks with a warm British accent (Alice voice). Toggle voice mode via the mic button.
Example: “I have diabetes, what happens with my application?”
Vera calls lookupKnowledge('endocrine') to understand diabetes in the context of insurance
Vera calls listConditions('Q2') to find all endocrine conditions in the rules engine
Vera calls lookupCondition('Q2', 'Q2_1_DIABETES') to get the full decision tree
Vera explains: which questions would be asked, what the possible outcomes are, and what factors matter (e.g. HbA1c level, medication, complications)
Every fact is traced to a verified source. Nothing is guessed.
Coming next
searchProviders
Find healthcare providers by location, speciality, and network
comparePlans
Side-by-side comparison of insurance plans based on user needs and conditions
fileClaim
Guide users through the claims process with document collection and status tracking
06
The rules Vera lives by
The chat is not a feature of Kindant. It is the entire experience. Every interaction, from researching cover to filing a claim, happens through a natural conversation with Vera. No forms, no dashboards, no buried settings.
Insurance is intimidating. Kindant should feel like talking to a knowledgeable friend, not filling out a medical form. Emerald greens, warm stone tones, rounded edges, generous spacing. Think Headspace meets Monzo.
Every recommendation shows its reasoning. Underwriting decisions cite the specific rules that were applied. When Vera does not know something, she says so. No black-box moments.
Underwriting decisions are made by a deterministic rules engine, not AI. The decision tree Michael built is converted into structured data that a system walks step by step. Every path is testable, auditable, and explainable.
80%+ of HK internet use is mobile. Every component is designed for a thumb first, mouse second. Bottom-anchored input, swipeable cards, voice as a first-class input method.
Start simple: 'What are you looking for?' End sophisticated: deterministic underwriting, plan comparison, claims filing. The interface grows with the user's needs, never front-loads complexity.
07
What has been shipped so far
12+ sessions, from project foundation to production handover. Every feature, every decision, tracked chronologically.
08
Ready for handover, 22 May 2026
09
Confirmed across kickoff and build
UK first, Hong Kong ready
UK primary market. Bilingual support (EN + zh-HK Traditional Chinese) shipped. Underwriting data already bilingual.
Invite-only closed beta
Working moms and busy professionals with unused policies. 20 invite codes ready. Google SSO also available.
Companion, not comparison
Kindant is a navigator and guide, not a comparison engine. No insurance sales, no kickbacks. Insurer-agnostic.
Human-in-the-loop is the USP
AI handles 90-95%. Sarah and Michael handle nuanced cases via the admin escalation dashboard.
All 4 journeys live
Guided Choice, Living Protection, Trusted Guide, and Transparent Resolution. Voice, image upload, PDF analysis all active.
250 automated tests
Deterministic underwriting engine validated against every condition path in all 15 disease categories. 17 agent scenarios pass.