Success stories
Zero to MVP, one voice at a time
Building an AI reading tutor for users who can’t read the instructions — from unproven prototype to a monetizable, classroom-ready product.
The client
EdTech company building a personalized, voice-based reading tutor for children 8–13 who aren’t yet fluent readers — many with dyslexia or ADHD.
Challenge
Move from a fragile prototype to a platform able to charge real families and run in classrooms — without an in-house platform team: a vendor-agnostic engine, phoneme-level scoring instead of right/wrong, and safe dev without risking children’s data.
SERVICE
AI Engineer — Backend, AI Engine & DevOps
INDUSTRY
EdTech · Generative & Voice AI · SaaS
TECHNOLOGY
FastAPI, LangGraph, Redis, AWS (Terraform, RDS, S3), Supabase, SpeechAce, ElevenLabs, React
KEY TECHNICAL DECISIONS
Adapters before vendors.
Swapping a provider means changing an adapter, not the engine.
Instrument before optimizing.
Full attempt logging shipped first, surfacing a hidden scoring gap.
Fail safe, not strict.
Scoring skips verification under ambiguity rather than penalizing the child.
Solutions
FastAPI backend + LangGraph state machine (listen → transcribe → classify → score → decide → respond), Redis perturn state
Provider-abstraction layer for STT, scoring, TTS and LLM, with zero-cost mocks for full-engine testing
End-to-end attempt logging — audio, transcript, scores, state snapshots — migrated to Supabase
Centralized rubrics plus the “taught grapheme” rule: only the phoneme being taught counts as a pass
Isolated dev/staging/prod on AWS, schema-as-code, CIenforced git flow
Infrastructure as code on AWS (Terraform) plus CI/CD with typechecking and automated tests
VAD spike, benchmarking a neural detector against real production recordings
Benefits
- Vendor-agnostic AI engine, swappable via adapters
- Full attempt-level observability replacing guesswork with production data
- Phoneme-level scoring that protects the core pedagogy
- Three isolated environments, schema-as-code
- Production-ready platform — unlocked Stripe monetization and real classroom use
What the client says
“We could not have reached this stage of the project without the efforts of the
Mindtech team, and for that we are grateful for our continued relationship and your support of our software build out.”
Graham Hillis
Director of Software – Vault
Benefits
Improving product descriptions on the online store to enhance the customer shopping experience and increase sales.
Challenge
Solution
The project is based on the use of AI generative models, specifically Gemini Pro Vision, to achieve three main objectives:
- Create new product descriptions by extracting attributes from product images.
- Enrich these product descriptions using previous information from other products
- Perform a retroactive process to update the product descriptions when new attributes are found when compared to embeddings generated by Gecko (Google model), i.e., add this information to existing similar products.
