mier.dev
All projects
2024LiveBachelor Thesis — analysis of the MTF system, voice control & AI summary implementation

MTF — Mein Technik-Finder

Accessibility-focused healthcare product recommender for the Berlin Senate, extended with voice control and AI-generated summaries

Mein Technik-Finder (MTF) is a questionnaire-based recommender system for Pflege 4.0 healthcare and assistive-care products, commissioned by the Berlin Senate. Users answer a guided questionnaire about their care situation and receive personalized product recommendations plus a structured PDF report they can take to a consultation. This bachelor thesis at TU Berlin's DAI-Labor analyzed the full MTF stack — questionnaire flow, recommendation logic, backend and product database — and then extended it with two AI-powered accessibility features: hands-free voice control and an LLM-generated summary of the user's answers and wishlist embedded into the final PDF.

View details Live site
2
AI features shipped
1
Bachelor thesis (TU Berlin)
PDF
Structured report output
iFrame
Embeddable on partner sites
01

Overview

MTF is a healthcare product recommender for the Pflege 4.0 market, designed for users who struggle to navigate the fast-moving care-tech landscape on their own — especially elderly users and people with impairments.

Users move through a multi-stage questionnaire describing care problems (e.g. falls, disorientation, hydration). Each answer narrows down which product categories are relevant, and the system surfaces matching products with filterable attributes like WLAN-free, no recurring costs, TÜV-tested or interoperable.

Selected products are collected in a wishlist and exported as a structured PDF that the user can take to a consultation appointment.

The thesis added two new AI features on top of the existing system: voice control for hands-free navigation, and an automatically generated natural-language summary of the user's answers and chosen products that gets embedded into the PDF.

02

Architecture

User → MTF web UI (embedded via iFrame) → Questionnaire & wishlist → Backend + product DB → AI services (voice + summary) → PDF report

Client (host site)
Embedded via iFrame on partner pages
L1
Guided questionnaireFilterable resultsWishlistVoice controlPDF export
Recommendation logic
Driven by explicit user answers
L2
Stage-based questionnaire flowCare-problem → category mappingAttribute filters (WLAN, cost, TÜV, etc.)Transparent, non-tracking ranking
Backend & product database
Care-tech catalog with rich attributes
L3
Products & categoriesCare problems → product mappingAttributes: interoperability, TÜV, cost, …
AI services
Voice control & summary generation
L4
Speech recognition → command parsingGPT action-matching from allowed actionsGPT summary over answers + wishlistSummary injected into final PDF
03

Tech stack

Frontend
JavaScriptWeb UIiFrame embed
Backend
Relational product databaseREST APIsPDF generation
AI
GPT for voice command matchingGPT for summary generationSpeech recognition / NLP
Context
TU Berlin — DAI-LaborBerlin Senate-funded projectBachelor thesis
05

MVP features

Guided questionnaire

A staged questionnaire walks the user from broad care topics (e.g. falls, disorientation) down to concrete care problems and matching product categories.

Transparent recommendations

Recommendations are derived from the user's explicit answers, not passive tracking — keeping the logic understandable and auditable for users and consultants.

Rich product attributes

Products are tagged with attributes like WLAN-free operation, no recurring costs, TÜV certification and interoperability so users can filter by what actually matters to them.

Wishlist & structured PDF

Chosen products are collected in a wishlist and exported as a clean, structured PDF the user can bring to a real consultation appointment.

Voice control

A GPT-backed voice control layer maps spoken commands to the system's available actions, so users with mobility or vision impairments can navigate without a mouse or keyboard.

AI-generated summary

An LLM-generated summary turns the user's answers and wishlist items into a cohesive natural-language overview that's appended to the PDF, giving consultants instant context.

Inclusive embedding

Delivered as an iFrame embed so it can be added to existing partner sites without disrupting their layout or design system.

06

Workflows

Recommendation flow

  1. 1

    User starts the questionnaire embedded in a partner site and selects the broad care topic that applies to their situation.

  2. 2

    Stage-by-stage follow-up questions narrow down to specific care problems and the product categories that address them.

  3. 3

    Matching products are surfaced with filters for attributes like WLAN-free, no recurring costs, TÜV certification and interoperability.

  4. 4

    Selected products are added to the wishlist and exported as a structured PDF the user can take to a consultation.

Voice control

  1. 1

    User activates voice mode and speaks a command in natural language.

  2. 2

    Speech is transcribed and forwarded to a GPT model together with the list of currently allowed actions.

  3. 3

    The model picks the best-matching action from that allow-list, keeping voice control scoped to what the UI actually supports.

  4. 4

    The chosen action is executed in the UI — moving through the questionnaire, toggling a product in the wishlist, or exporting the PDF — entirely hands-free.

AI summary generation

  1. 1

    When the user finishes the questionnaire, their answers and wishlist items are collected as structured input.

  2. 2

    That input is sent to a GPT model with a prompt asking for a cohesive, plain-language summary of the user's situation and chosen products.

  3. 3

    The generated summary is injected as a dedicated section of the final PDF report, alongside the structured wishlist.