All Projects

First Engineering Hire:ServioAI → Pairio

VerticalAIforindustrialmachinemaintenance—firstengineeringhireona3-personteam,nowPairio
3-Month Sprint · First Engineering Hire
Y Combinator-backed after my departureIndustrial MaintenanceProduction RAGOct–Dec 2025

The Project

Joined ServioAI as the first engineering hire alongside the two founders — Tim Zinkl and Matthias Wolf. A 3-person team building a vertical AI platform for industrial machine maintenance: FastAPI backend, Next.js web app, Expo mobile app for field technicians, and a Microsoft Office add-in for documentation workflows. Worked full-stack across all four surfaces from October to December 2025, with a soft rule: be in the customer conversations. In vertical AI, the model is only ever as good as the workflow you wired it into.

Left at the beginning of January 2026 to focus on my master's thesis at TUM. After my departure, the product rebranded as Pairio, which now presents itself publicly as Y Combinator-backed.

4 platforms

web · mobile · Office · dashboard

FastAPI + Next.js

TypeScript end-to-end

Vertex AI

Google + Supabase

SSE streaming

real-time chat & docs

System Architecture

One Monorepo · Four Surfaces · One AI Core

Web
Next.js
Analytics
Internal
FastAPI
Vertex AI
Mobile
Expo
Office
Add-in

FastAPI core · Vertex AI services · Supabase / self-hosted SQL · SSE streaming end-to-end

Featured Wins

Production RAG

Smart Context Selection

Before: users had to manually pick which documents went into the AI's context, then we'd dump up to 300 pages of raw text into every request. Slow, expensive, and fragile when documents grew.

After: the user just asks the question. The system identifies the relevant documents, pulls only the chapters that matter, and stitches them into a context that is ~80% smaller without losing any retrievable information. Built as a generalisable context tool used across analyses, chats, and documentation flows.

~80% smaller · lossless

~80% context window reduction · zero info loss

Backend Evolution

MongoDB → Self-Hosted SQL

The product started on MongoDB and outgrew it: relational queries got expensive, and document-shaped data hid real relationships we needed to enforce. We moved to Supabase for the speed of a managed Postgres, then to a self-hosted SQL setup once data residency and cost control mattered more.

Both migrations happened on a running product — schema changes, dual-writes, deletion sync, and a Supabase sync script kept the existing customer-facing flows running while the new tables took over.

zero downtime · faster on the queries that matter

Two migrations while customer-facing flows stayed live

Vertical AI

Video → SOP Pipeline

Field technicians capture maintenance work on Meta smart glasses. The pipeline takes that raw video and turns it into a structured Standard Operating Procedure — extracted steps, timestamped photos, German-language formatting, ready for editing in a custom in-app SOP editor.

Iterated through three versions, moved photo storage from inline blobs to Cloudflare R2, added PDF generation with the customer's mocked header layout, and rebuilt the editor UI for review-and-publish workflows.

video timeline → structured SOP

Field video → reviewable SOPs in minutes, not hours

Speed of Execution

Internal LLM Analytics Dashboard

Spun up a separate Next.js dashboard so the leadership team could see what was happening inside the AI: token usage, model latency, error rates, SOP generation throughput, and per-organisation analytics. Auth gated to the analytics team via PropelAuth, direct queries against the production Supabase.

First production version shipped in one week — initial push to deployed dashboard with home metrics, SOP analytics, and documentation views. Polished into v2 the following week.

7 days · idea → live

Idea → production tool in 7 days

Final Month

Web → Mobile Port

With the web stack stable, the priority for December 2025 became putting the same capabilities into the hands of field technicians directly. Took the entire web feature set — smart context, chat, document workflows, SSE streaming — and ported it to the Expo mobile app.

Shared TypeScript types between web and mobile, reused the FastAPI endpoints unchanged, kept the streaming UX consistent across surfaces. The app left my hands ready for the field.

feature parity · web → mobile, Dec 2025

Full feature parity ported in 1 month

Tech Stack

FastAPINext.jsTypeScriptTailwindShadcn UISupabase / PostgresVertex AIPropelAuthExpo (React Native)React + Vite (Office Add-in)Cloudflare R2GitHub Actions CI/CD