Skip to content
Open to work
Case study

Private · NDASep 2025 – Aug 2026

Commercial AI SaaS platform

An AI-powered SaaS platform for content analysis, built from first commit to production.

Under NDA: no links, screens or names. Everything here is worded as in my CV.

Case study · The brief

A US-based company needed a commercial AI-powered SaaS platform for content analysis — from the first commit to production. I designed the architecture and built the platform end to end: a Python / FastAPI backend, a Next.js / TypeScript frontend with a shared design system, Google Cloud infrastructure, CI/CD, Stripe billing and a multi-agent LLM pipeline.

What I built

  • Multi-agent LLM pipeline: specialised agents run in parallel and their structured outputs merge into one scored report.
  • Fault tolerance for non-deterministic AI output: incomplete results never reach users, and failed runs recover without repeating costly steps.
  • MCP server with secure OAuth 2.0 authorization, letting AI assistants work with the product on users’ behalf.
  • 1,300+ backend tests, 80+ documentation pages including 20+ ADRs, 1,100+ commits.
My role
AI-Native Fullstack Engineer, contract with a US-based company
Period
Sep 2025 – Aug 2026
Stack
  • Python
  • FastAPI
  • MongoDB
  • LLM APIs
  • MCP
  • TypeScript
  • Next.js
  • React
  • Tailwind CSS
  • Docker
  • Google Cloud

How it fits together

Drawn from my CV and worded as it is — no more detail than that. The outline shows who built each part.

  • Web appNext.js · TypeScript — Built by me
  • BackendPython · FastAPI — Built by me
  • Specialised agentsworking in parallel — Built by me
  • Scored reportnever shown incomplete — Built by me
  • MCP serverfor AI assistants — Built by me
  • AI assistants — People and outside services

The same, step by step

  1. Web app (Next.js · TypeScript) with Backend (Python · FastAPI), both wayscontent · progress
  2. Backend (Python · FastAPI) to Specialised agents (working in parallel)parallel runs
  3. Specialised agents (working in parallel) to Scored report (never shown incomplete)structured outputs
  4. AI assistants to MCP server (for AI assistants)OAuth 2.0
  5. MCP server (for AI assistants) to Backend (Python · FastAPI)on users’ behalf
Built by mePeople and outside services

Key decisions

The calls that shaped the system, and why they were made.

  1. D1

    Specialists in parallel, one report

    Specialised agents work in parallel, and their structured outputs are combined into a single scored report.

  2. D2

    AI output is not trusted to be complete

    Fault tolerance designed for non-deterministic AI output: incomplete results never reach users, and failed runs recover without repeating costly steps.

  3. D3

    Stable under concurrent jobs

    Resource usage optimised for heavy processing workloads, keeping the service stable under concurrent jobs.

  4. D4

    AI assistants as first-class clients

    An MCP server with secure OAuth 2.0 authorization lets AI assistants work with the product on users’ behalf.

  5. D5

    One gate for people and agents

    Development moved to an AI-native workflow: custom subagents for research, testing, review and documentation checks, backed by one automated quality gate — linting, typing, security scans, architecture rules, test coverage — shared by local runs and CI.

Outcome

backend tests
1,300+
documentation pages
80+
ADRs
20+
commits
1,100+
  • In production with subscription billing and usage quotas, real-time progress updates and an admin console.
  • Privacy features: data export, account deletion and data retention.
  • Automated CI/CD with secure deployments, safe database migrations and isolated environments.