Skip to the case study
Case study01 · OverviewTuring

AI slide generation that writes in each teacher’s own voice.

TeacherAIde learns a teacher’s style from their own lecture materials, then generates decks that already sound like them - designed, built and shipped by a single engineer.

IndustryEducation technology
RoleSole engineer - Full Stack & AI
Core AIAnthropic Claude

Voice-matched decks

Style fingerprint, scored per deck

Billing built in

Stripe plans, webhook-verified

One engineer

Backend, frontend and AI workflows

Case study02 · Solution & capabilities2 / 4
02
The solution

A first draft that sounds like you.

Generic AI tools give teachers slides in the wrong structure, tone and depth. TeacherAIde measures how a teacher teaches, then generates decks that match.

1Upload lecture materialsThe decks and notes they already have
2Style is extractedWeighted structural and narrative metrics
3Generate and exportDeck lands in Google Slides

Key capabilities

What the product does

Style fingerprint

Learns how a teacher teaches.

  • Weighted style metrics from uploads
  • Scores new decks for consistency

Generate & export

From fingerprint to a deck ready to present.

  • Outline, content, personalization, notes
  • Google Slides export, PDF and preview

Plans & billing

A paid product from the first day.

  • Stripe subscriptions and plan changes
  • Server-verified webhooks drive account state

Admin dashboard

Platform health at a glance.

  • Generation runs and their status
  • Teacher accounts and billing health
Case study03 · Architecture & build3 / 4
03
Architecture

Structured by design. Tuned on real data.

Next.js · NestJS · AWS Lambda · PostgreSQL · Claude · n8n · Stripe

Architecture diagram: a teacher uses a Next.js web app and NestJS API on AWS Lambda; n8n workflows call Anthropic Claude in stages; the system uses Amazon S3 for direct pre-signed uploads, PostgreSQL for users, runs and styles, Stripe for subscriptions, Google Slides for OAuth export and Google Gemini for image rendering only.

How it was built

Four steps to production
1DiscoverStyle as weighted metrics2ArchitectPrompts, schemas, tool calls3BuildPipeline, export, billing4LaunchCI/CD to AWS and n8n

Challenges

Problem, then fix
Prompts edited by hand in n8nEach prompt is a versioned file, injected at publish time and deployed by CI.
Cost and drift at scalePrompt caching tuned against measured dev/UAT runs, not assumed.
Free-text model outputEvery call answers through one named tool call, so the backend gets structured data.
Case study04 · Results & impact4 / 4
04
Results

What it delivered.

11AI workflows, every prompt version-controlled
4generation stages, outline to speaker notes
2AI models, cost and failure isolated per stage
1engineer across the whole stack

Impact

Who it changed things for
For teachers

A draft that sounds like them

Decks arrive in the teacher’s own structure and tone, not a template to rewrite.

For the business

Billing you can trust

Stripe subscriptions and plan changes, with server-verified webhooks driving account state.

For the team

Changes ship like code

Prompts and pipelines are versioned and deployed by CI, with no manual n8n export step.

One engineer shipped a production AI product - schema and Stripe billing, Google Slides export, prompt design and CI/CD.

Muhammad Abdullah · AI Engineer, TuringStart a similar project