HTI.560 Conversational AI · Tampere University
A multimodal conversational AI that helps patients disclose sensitive health information before a doctor's appointment — without the fear of human judgment.
Live Deployment · GitHub Pages
The app is hosted and fully navigable. The AI screening chat requires a Groq API key — not embedded for security reasons. To try it: paste your key in the app's Settings, or contact Garv for a live walkthrough.
App Screens
Real screens from the deployed application — scroll to explore
Home
Chat — Start
Chat — Input
Voice — Thinking
Voice — Listening
Mid Session
Review Summary
Emergency
Settings
Past Summaries
Patients lie to their doctors. Not out of malice — out of shame.
"Patients are often more willing to disclose stigmatized information to a machine because it lacks the capacity to judge socially." — Blease, The BMJ, 2024
Sensitive symptoms — substance use, sexual health, mental health struggles — are routinely withheld or downplayed during clinical appointments. Patients fear moral judgment from doctors. Clinical time is limited; a 15-minute slot rarely builds the trust needed for honest disclosure.
This information gap leads to inaccurate diagnoses and worse care outcomes. The question was: could a carefully designed AI agent, positioned before the appointment, fix this?
We grounded every design decision in academic literature on the Automaton Effect, ethical AI in healthcare, and patient-chatbot interaction styles.
The Automaton Effect
Research by Blease (2024) confirms patients disclose more sensitive information to machines than to clinicians, precisely because the machine cannot socially judge them.
Fake Empathy Backfires
CHI 2024 research by Lyu et al. showed that chatbots saying "I'm so sorry for your pain" were perceived as creepy and insincere. Users preferred "Clinical Validation" — clear, task-focused acknowledgment.
Autonomy is Non-Negotiable
Ethical AI literature on mental health tools (JMIR 2025) found that explicit opt-out controls are critical — users need to feel in control, not interrogated.
Users Are Emotionally Vulnerable
Clinical environments involve anxiety, pain, and distraction. Interfaces must be high-contrast, forgiving of errors, and never force a yes/no binary on ambiguous symptoms.
Anxious Alex
Age 22 · Seeking advice on sexual health
Pain Points
— Fear of moral judgment
— Intimidated by authority figures
Goals
— Disclose symptoms accurately
— Feel reassured of confidentiality
Private Patricia
Age 65 · Experiencing early signs of cognitive decline
Pain Points
— Deep embarrassment
— Fear of losing independence
Goals
— Slow-paced, patient interaction
— Explain symptoms in her own words
01 — Literature Review & Concept
Grounded the entire system in the Automaton Effect and established the ethical opt-out framework. Defined personas and scoped the system to pre-clinical screening — not diagnosis. The AI was never allowed to act like a doctor.
02 — Conversation Scripting
Drafted dialogue trees for high-sensitivity conditions (anxiety, sexual health). Applied the "Clinical Validation" principle from CHI research — acknowledgment over fake empathy. Wrote the two-step closing protocol to ensure the AI always summarizes before ending.
03 — Interface Design in Figma
Designed the full component library — the Voice Orb states, quick-reply button variants, the sticky action bar with persistent "Prefer not to answer" affordance, and the final PDF Review screen. High-contrast, forgiving, and mobile-first throughout.
04 — AI Integration & Prompt Engineering
Connected the front-end to the Groq API (Llama-4-Scout) using strict JSON-mode prompting. This constrained the AI to ask exactly one question per turn, prevented unsolicited diagnoses, and reliably generated dynamic on-screen buttons alongside spoken responses.
05 — Voice UX & Error Handling
Implemented the Web Speech API with custom pitch-shifting logic for three distinct voice personas. Designed multimodal fallback: if speech recognition confidence dropped below 60%, the system automatically switched to on-screen buttons, preventing dead ends.
06 — The Handoff Redesign
The original plan emailed the summary to the doctor — until we asked one question: does any patient actually know their doctor's email address? Nobody does. We replaced the entire flow with a QR code: the summary is compressed into a link the doctor scans at the appointment, opening a clinical document on their own device. More realistic, and more private — the data never touches a server.
A real exchange from the anxiety pre-screening flow. Notice the agent never forces a yes/no — it always offers an escape hatch.
Design note — The agent never says "I'm sorry you're going through this." It says "I've noted that." Clinical validation over performed empathy. CHI 2024 research showed the neutral bot was consistently rated more trustworthy.
A zero-backend, multimodal conversational web app. Every design decision was driven by one constraint: the patient must always feel in control, never interrogated.
Neutral-Supportive Persona
Clinical validation over fake empathy. The agent acknowledges inputs clearly without performing emotions it cannot actually have.
Explicit Autonomy
A persistent "Prefer not to answer" button is always visible. Every sensitive topic opens with a reminder that skipping is always an option.
No Dead Ends
Speech recognition failure, hesitation, or ambiguity never ends the conversation — the system always offers a clear next step.
Voice Orb with Silence Detection
A continuous listening interface with distinct visual states — idle, listening, processing. Silence detection automatically stops recording, removing the need to tap a button.
Dynamic Quick-Reply Buttons
The Groq LLM generates context-aware on-screen buttons for each question in real time via JSON prompting. Users can tap or speak — both inputs are equally first-class.
Two-Step Closing Protocol
Before any session ends, the AI is forced to produce a structured summary and ask for the patient's final confirmation. Nothing gets submitted without explicit patient approval.
QR Clinical Handoff
The summary is compressed into the URL fragment of a shareable link — the doctor scans a QR code and a clinical document opens on their own device. Fragments never reach a server, so patient data never leaves the patient's control.
AI Engine
Groq API (Llama-4-Scout)
Strict JSON-mode prompting. Constrained to one question per turn. Prevented from diagnosing.
Voice Engine
Web Speech API + ElevenLabs
Custom pitch-shifting logic for Male, Female, and Neutral voice personas. Multimodal fallback at <60% confidence.
Frontend
HTML / CSS / Vanilla JS
Deployed live via GitHub Pages. No framework dependencies — fast load, zero build overhead.
Privacy Layer
localStorage + jsPDF
All session data stays on-device with enforced 30-day auto-deletion. Clinical PDF generated in-browser. No backend database.
Clinical Handoff
QR + Web Share + lz-string
Summary compressed into a URL fragment — fragments never touch a server. The doctor scans a QR and gets a print-ready clinical document instantly.
Design
Figma
Full component library covering all interface states — orb, buttons, summaries, error states.
UX Engineer & System Architect
System architecture — zero-backend privacy model
QR clinical handoff design (lz-string URL encoding)
Groq API integration and JSON-mode prompt engineering
Constraining LLM to clinical-safe behaviour (no diagnosis)
Multimodal voice state machine (orb, silence detection)
UI localization across 10 languages, synced with Whisper + LLM
Error handling for API rate limits with animated fallback UI
GitHub deployment pipeline and live app delivery
Figma component library for interaction states
What I Learned
Bridging UI design and live AI constraints is its own discipline. Writing prompts that produce predictable, safe, one-question-at-a-time clinical dialogue required weeks of iteration — not minutes.
We underestimated the complexity of voice-to-text mode switching. State management for overlapping audio streams had to be fully rewritten mid-project to fix an audio collision bug.
Designing for vulnerable users sharpens every instinct. Error messages, loading states, and fallback paths cannot be afterthoughts when the user might be anxious, in pain, or distracted.
15–20
Questions per session
Strict LLM prompting kept the screening focused and clinically efficient.
10 min
Live demo validated
Fully functional prototype demonstrated in real time to course faculty.
0
Backend servers needed
Complete zero-backend privacy architecture — patient data never leaves the device.
The most significant design outcome: the AI was successfully constrained to never act as a doctor. It asks. It records. It hands off. That boundary — between data collection and clinical judgement — was the hardest design problem and the most important one to get right.
The app is live and functional. These are the concrete next milestones planned for the clinical handoff phase.
Platform
ShippedProgressive Web App
Service worker + manifest so patients can install Talos directly to their phone home screen. Offline caching keeps the UI functional without internet.
Platform
ShippedMultilingual Support
Expanded from 3 to 10 languages. Whisper voice transcription, LLM responses, and the full UI all sync to the selected language.
Clinical
ShippedPre-Screening Intake Form
A dedicated optional form before the chat collects age, medications, and allergies. Frees the AI to focus purely on symptom screening.
Clinical
ShippedQR Clinical Handoff
The summary travels inside a QR code / shareable link as a compressed URL fragment — the doctor scans it and a clinical document opens on their device. Replaced the original email plan: patients don't know their doctor's email.
Resilience
ShippedRate-Limit Error Handling
HTTP 429 responses trigger an animated banner with a live countdown timer and auto-retry logic — the user never hits a dead end due to API throttling.
Evaluation
A/B Usability Study
Comparative test between Talos and a standard digital intake form, measuring "Ease of Disclosure" scores with real volunteers.
Talos Care sits at the intersection of conversational UX, clinical ethics, and real AI implementation. Most portfolio projects exist as Figma prototypes. This one runs in a browser, speaks ten languages in a chosen voice, and hands the doctor a scannable clinical document — no server ever sees the patient's data.
Designing for patients who are afraid required a level of care I hadn't applied to consumer apps. Every label, every fallback message, every moment of silence had to be intentional. The research wasn't decoration — it was load-bearing. The "Neutral bot is more trustworthy than Empathetic bot" finding directly rewrote our scripting.
What I carry forward: the habit of designing the unhappy path with the same rigor as the happy path. And the understanding that in high-stakes interaction design, the system's limits are as important as its capabilities.
Open to UX/UI internships — Finland & Europe
Interested in working together?