Andreas Kater

I re-platformed a legacy Servoy-based ERP for a seafood wholesaler onto a modern cloud stack. I developed a React front end using Vite and React Query, featuring dashboards and a kiosk mode. The back end was built with a Node.js API on Google Cloud Run, and the database was migrated to PostgreSQL on Cloud SQL. For security, I implemented Firebase Auth and Secret Manager.

I successfully migrated production data from the legacy system and introduced an AI-powered ordering assistant using OpenAI and Twilio, allowing customers to place orders via WhatsApp. This transition was accomplished while ensuring the legacy system remained fully operational for daily use.

I developed the back end of Aurélie, an AI phone-receptionist SaaS tailored for small businesses. The system handles real phone calls using Twilio Media Streams and bridges live audio over WebSockets to OpenAI's real-time API with voices from ElevenLabs/Cartesia. This setup enables the AI to take bookings, answer questions from a per-business knowledge base, and smoothly hand off to humans through live and virtual call queues with hold and conference control.

I designed the PostgreSQL data model using Knex/Objection on Cloud SQL and created a multi-organization REST API with Swagger documentation. Additionally, I implemented an MCP tool layer for the AI agent, post-call transcription and analysis, SMS workflows, and the complete commercial stack: Stripe checkout, subscriptions, billing portal, and self-serve onboarding. The entire system is deployed on Google Cloud using Docker and GitHub Actions for CI/CD.

I am leading the modernization of PMQA, a multi-tenant SaaS for quality assurance and auditing used by UK local-authority Benefits & Revenues teams. The project involves transforming a 15-year-old Zend/PHP codebase into a more modern stack using React 18 and TypeScript for the front end, along with a Node.js 20, Express, and TypeScript API. We're employing a strangler-fig approach, allowing the new system to read and write to the existing legacy MySQL schema and operate alongside the old application until each council signs off on the transition.

I designed the layered architecture, incorporating routes, controllers, services, and data management, all strictly using TypeScript. The platform supports database-per-council multi-tenancy, XLSX/PDF reporting, and a database-backed email queue. The deployment strategy includes Docker Compose for local setups and Cloud Run and Cloud SQL on GCP, with Cloud Build CI enabling automated accessibility checks.

The migration is structured workstream-by-workstream with comprehensive architecture and cutover documentation. Over 120 automated test files ensure behavior parity with the legacy system, maintaining functionality and minimizing disruptions during the transition.

I built a comprehensive self-hosted UK payroll application that's currently used in production. At its core, it features a unit-tested pure-JavaScript tax engine that implements the HMRC's 2026-27 rules. This includes PAYE income tax for rUK and Scottish bands, covering every tax-code variant (L, BR, D0/D1, NT, 0T, K codes, W1/M1), Class 1 National Insurance for nine category letters (both employee and employer), three pension tax-relief mechanisms, all five student-loan plans, and statutory payments like SSP, SMP, and SPP.

The React/Vite front end facilitates employee management, payroll runs with cumulative year-to-date tracking, and generates one-click A4 PDF documents such as payslips, P60s, P45s, P30s, as well as FPS/EPS RTI XML files compliant with HMRC schema. Prioritizing privacy, the system ensures that all payroll data is stored in browser localStorage, with JSON export and backup options, so no personally identifiable information (PII) touches the server. The application is deployed on Google Cloud Run with CI/CD, and I served as the sole architect and developer for this project.

An advertising platform wanted to replace an old analytics platform with a modern React application that pulled information from their brand new API. Since the data sets were non-trivial and needed to be summarized into various views, the app made extensive use of redux and sagas. Some views required database queries that the API was not yet capable of, so I added endpoints to the Node.js Feathers back end and add that functionality to the API.

A luxury resort wanted to relaunch their website with a map view showing what is on offer on their premises. I rebuilt their website using React and Next.js and added the map view. This is powered by the GoogleMaps JavaScript API and allows clicking on buildings to get more information about them below.

Ery-API is a Node.js-based CLI tool to give you quick access to the API calls you're looking for and it performs the OAuth flow for authorization with all Google APIs.

The idea is to never have to read API Docs again. Give it a go with:
npm i ery-api -g

The business I was working for wanted to implement a 360-degree feedback platform, where no off-the-shelf solution did exactly what was required, so I built an app using Firebase and React that would allow users to pick the people they requested feedback from, collected this feedback, allowed managers to summarize it, and to ultimately set goals for each employee that were tracked over time; GCP and Google Cloud functions are used. Thanks to Firebase's real-time DB, comments were shared instantly across open tabs, which was important for the quick completion of the feedback cycle.

For a business I worked for, I created a Dialogflow based HR chatbot that allowed our team to get quick answers to questions like: "How do I submit an expense?", "How many holidays do I have left?", "What is our policy on XXX?", etc. I integrated it into a slackbot to give everyone in the business fast answers to questions on how we do things.

I built a server that allows you to snooze emails, but forwarding an email to, for example, "3hours@snoozethat.com" or "3nov3pm@snoozethat.com" and getting a reply in the same email chain at the specified time. This helped our customer service team stay on top of customer accounts when delayed actions/follow-ups were required (This was before Gmail introduced the snooze function).

I built a platform that helps you learn a language by translating an epub e-book word-for-word. This allows you to read a book in a language you are learning and always having the translation available if you don't know a word in it. The tool will also pronounce the word for you if you tap it and show you further information, such as the conjugation if it is a verb. As this was a Progressive Web App, it allowed you to read the books when you were offline.

I built a tool that used a service worker to create a progressive web app that works offline. The tool was designed to help you keep track of your tasks by allowing you to snooze tasks until the time comes to take care of them.

I built an ERP tool for an art gallery that needed to track where certain pieces of art were stored, who had enquired/bid/purchased them, and what invoices/communications were logged against them. The tool also had a front-end component that acted as the gallery's online presence.

I designed and built a full-stack event management platform for a corporate client, featuring a React/Tailwind frontend and an Express API running on Firebase Cloud Functions. The back end is supported by PostgreSQL, managed through Prisma.

I set up the entire delivery pipeline using GitHub Actions for CI/CD, including automated database migrations. The development process included unit tests with Vitest and end-to-end tests with Playwright. I owned the project from the initial client brief and roadmap planning through to successful production deployment.

I architected and built the optimization engine for ShiftHarmony, a staff-scheduling product tailored for medical practices. This engine processes inputs such as a month, a roster of doctors, PAs, and nurses, shift definitions, vacation plans, and a rulebook to compute an optimal monthly schedule using Google OR-Tools CP-SAT constraint programming.

A standout feature is the natural-language layer, allowing practice managers to write rules like "Dr. Smith only works mornings" in plain English. An OpenAI-powered pipeline translates these rules into parameterized constraint templates, ensuring they are validated before reaching the solver. The system differentiates between hard rules and soft preferences, reports violations graded by severity, re-validates every solution after solving, and includes scenario-based tests and HTML schedule visualizations.

This comprehensive solution is built in Python with Pydantic data models. I owned the full design, including the constraint model, the LLM rule-generation pipeline, and the solver integration.

I architected and led the development of the back end for a multi-tenant SaaS invoicing platform. The core is an Express/PostgreSQL API with over 40 migrations that manage the entire invoicing lifecycle, from creation to scheduling and reminders. It uses Puppeteer to generate PDFs through a managed browser pool and handles templated email delivery.

I integrated two payment rails: Stripe, including Stripe Connect, and GoCardless for direct debit transactions. Additionally, I implemented TrueLayer for open banking, enabling bank-account connectivity and automatic transaction reconciliation, along with Wise for transfers.

The platform facilitates bidirectional sync of invoices and payments to Xero and QuickBooks using authenticated webhooks and hash-based change detection. A Slack integration also pushes payment events to users' workspaces. It supports EU, UK, and US tax handling, subscription billing with trials, Firebase authentication, Google Cloud Storage, and Cloud Tasks. The system is deployed through separate staging and production GitHub Actions pipelines on Google Cloud.

I founded and built Veqaro, a SaaS platform designed to automate the collection of overdue invoices for small businesses. I was responsible for designing and developing the entire product, encompassing a public marketing site with pricing and legal pages, and a multi-organization customer dashboard for managing users, invoices, and payments. This includes an in-browser PDF invoice preview and presigned uploads to Google Cloud Storage. I also developed a documented REST API with Swagger schema validation.

The platform's stack consists of a React 19 front end using React Router 7, Tailwind, and shadcn, hosted on Firebase Hosting with Firebase Auth. The back end is powered by a Node.js/Express server on Google Cloud, utilizing PostgreSQL (Cloud SQL) via Knex. The project is organized as an npm-workspaces monorepo, allowing for shared validation code between the frontend and backend.

Additionally, I built a companion automation service that manages outbound debtor communications across email, SMS, and payment links, further streamlining the invoice collection process.

I designed and built Rowmance, a multi-tenant SaaS tool for querying PostgreSQL databases from the browser, serving as a simple alternative to desktop tools like DataGrip. Users can sign in with Google using Firebase Auth, register database connections, write SQL with autocomplete in a CodeMirror editor, visualize schemas as interactive graphs, export results to Excel, and browse their last 100 queries with stored results.

I developed the React and TypeScript front end, a Node/Express API hosted on Google Cloud Run, and utilized Firestore for the data layer. Security was prioritized: connection passwords are encrypted with AES-256-GCM using a key stored in GCP Secret Manager, and Firestore access is restricted with deny-all rules, ensuring every read/write operation is authenticated through the API.

I set up CI/CD pipelines via GitHub Actions, which build the container with Cloud Build, deploy it to Cloud Run, and ship the SPA to Firebase Hosting with every push to the main branch. I served as the sole engineer, overseeing product development, architecture, security, and infrastructure.

I designed and built an AI-powered call-center platform for a US auto-repair chain. The platform features an AI voice agent, leveraging OpenAI and ElevenLabs with Twilio Voice, to handle inbound calls. It recognizes returning customers, answers service questions, and books workshop appointments at the correct branch through tool calls exposed via an MCP server.

I developed the entire stack solo, including Express/TwiML for call routing, managing live and virtual queues, conferencing, and hold music. Additionally, I created a React-based softphone dashboard for human agents and implemented a PostgreSQL data layer. For quality assurance, I set up an ffmpeg pipeline that mixes and analyzes call recordings with AI. The deployment was executed on Google Cloud/Firebase with CI/CD.

The system efficiently handles real customer calls from end to end and seamlessly transitions to human agents when necessary.

I developed a comprehensive ordering platform that enables restaurants to order from a seafood distributor via four channels, all integrated with the distributor's ERP system. The platform features an Expo/React Native app (iOS/Android/web) with Firebase authentication, offering a live catalog, cart, order history, and an AI chat agent. The chat agent, built using Gemini on Vertex AI, facilitates product search and order placement, and supports voice input via Google Speech-to-Text.

For phone orders, I implemented a Twilio media-stream WebSocket bridge to the OpenAI Realtime API, enabling callers to interact with an AI agent that identifies the customer, builds the order using fuzzy product matching, reads it back for confirmation, and saves it to the ERP. Email orders are processed via a Gmail watch and a Cloud Pub/Sub push pipeline, using OIDC verification and idempotent extraction.

The back-end infrastructure is built with Node.js/Express on Cloud Run, accompanied by a full test suite and CI/CD pipeline via GitHub Actions. The pipeline re-points Twilio webhooks after each deployment. I served as the sole engineer, overseeing everything from architecture to app-store rollout.

I engineered the v2 back end of Brainy RMS, a revenue-management system designed to help independent hotels optimize room pricing. This version was built as an event-driven monorepo using Firebase Cloud Functions. Connectors are used to download historical and live reservation data from the Octorate channel manager via OAuth and paginated APIs. Firestore triggers respond to reservation changes in real time, while pipelines compute occupancy, ADR, and revenue metrics in BigQuery for pricing analytics.

The system incorporates Cloud Tasks-authenticated functions and a shared internal npm package across services. Firestore security rules and local emulator suites with seeded sample data facilitate secure and efficient development. This project was delivered through my agency, Wakeflow.

I led the development of a workflow automation platform akin to Zapier. On the front end, I built a drag-and-drop canvas using Next.js 14, React Flow, and shadcn/ui, enabling users to create automations composed of trigger, conditional, delay, HTTP, data-transform, email, SMS, and WhatsApp nodes. The interface supports auto-layout, flow validation, and includes a Monaco editor for inline JSON and code.

On the back end, I developed the execution engine that processes these workflows. Each node type corresponds to a Node.js service, while Google Cloud Tasks manages scheduling and delayed steps. Variables are passed between steps using interpolation, with PostgreSQL (via Knex) storing definitions and runs. Pusher is used to stream live execution status back to the UI. The project was delivered with a small team through PR-based collaboration.

Orbio (orbio.one) is a modern CRM platform my agency designed and built from scratch; I architected and led development of its back end. The API is a Node.js/Express service backed by PostgreSQL (Knex migrations) with Socket.IO realtime updates, covering the full CRM surface: contacts, companies, deals and pipelines, conversations with Gmail-integrated email, meeting scheduling, invoicing with Stripe billing, e-signature workflows, web forms, segments, custom objects, a hiring module, LiveKit-powered video calls, and call recordings. It integrates Google Cloud Pub/Sub, Cloud Tasks, and Cloud Storage, Twilio, Slack, HubSpot, and Sage, and uses OpenAI/Anthropic models for AI features such as business-card data extraction. The API is documented with Swagger, validated with AJV JSON schemas, secured with Firebase Auth and JWT, and rate-limited, and shipped continuously via GitHub Actions.

I designed and built the back end for func.live, a platform that lets developers publish and invoke cloud functions via a single, consistent HTTP API with token-based authentication and usage-based billing. The Node.js/Express API (deployed on Google Cloud Functions with PostgreSQL on Cloud SQL) handles function registration and discovery, JSON Schema validation of inputs, proxying and logging of invocations, file uploads to Cloud Storage, and full Stripe integration: checkout sessions, signed webhooks, account balances, and per-call transactional charging.

I architected the system, built the core middleware stack (auth, logging, error handling), set up GitHub Actions CI/CD, and led a small team of developers contributing via pull requests. The platform is complemented by a Next.js web app, a published npm SDK, and a CLI that I also built.

I designed and built the tool-calling backbone of a ChatGPT-powered WhatsApp personal assistant. The service exposes about 50 self-describing functions the LLM invokes via OpenAI function calling, each publishing its own JSON-Schema docs endpoint. Capabilities include sending Gmail and creating Calendar events on the user's behalf (via a custom OAuth token broker), Google Drive and Contacts search, passport OCR, AI image editing and background replacement, translation via Google Cloud, Twilio SMS, QR and barcode generation, and a suite of data-format converters. Deployed to Google Cloud Functions with GitHub Actions CI/CD and Mocha test coverage. Built in mid-2023, weeks after function calling launched - an early production example of LLM agent tooling.

I designed and built an orchestrator that drives autonomous AI coding agents through a project backlog. The system schedules independent workstreams from a dependency graph and runs them concurrently in isolated git worktrees, each on its own branch, then serializes merges back to the base branch with automated verification (typecheck and tests) as a quality gate, with the agent resolving merge conflicts. Built in TypeScript on the Anthropic Claude Agent SDK, with mission files defining goals and per-lane scratch state kept out of version control. Demonstrates agent orchestration, scheduler design, and advanced git automation - infrastructure for supervising fleets of AI developers rather than writing code by hand.

I designed and built a full-stack deal-flow automation platform used to screen software businesses for acquisition. The system ingests Information Memorandum PDFs via drag-and-drop and an automated Gmail inbox watcher (Gmail API, Google Pub/Sub), stores documents in Google Cloud Storage, and persists structured deal data in Cloud SQL Postgres. An LLM-based rating agent scores every incoming deal against acquisition criteria on a daily Cloud Scheduler job, and a React dashboard surfaces a filterable pipeline with analytics.

I architected and implemented the entire product solo: TypeScript monorepo (Express API, React/Vite/Tailwind client), storage and database abstractions with full local/prod parity, Google Sign-In and API-key auth, Docker and Cloud Run deployment with GitHub Actions CI/CD, and a comprehensive automated test suite covering routes, migrations, Gmail ingestion, and the AI rating pipeline.

I designed and built a full cold email outreach platform on Google Cloud. The system connects Gmail and Microsoft 365 mailboxes via OAuth, runs multi-step sequences with configurable delays, and protects sender reputation with automated warm-up scheduling and quality scoring, per-domain send limits, mailbox rotation, send-time optimization, spam-placement testing with automatic rescue, and Google Postmaster metrics ingestion. It detects replies and bounces, tracks GDPR/PECR compliance, enriches contacts via Apollo, and uses the Anthropic Claude SDK to generate highly personalized messages.

I built the Node.js/Express API on Cloud Functions, the PostgreSQL schema and migrations, the React dashboard, the test suite, and the GitHub Actions deployment pipeline, plus complete operator setup documentation.

I built an AI-powered invoice ingestion service that replaced a paid SaaS subscription in my agency's accounting workflow. The Node.js/Express service receives invoice PDFs (including automatic ingestion from Gmail), uses the Claude API to determine whether each PDF is an invoice, classifies it as accounts payable or receivable, and extracts invoice number, dates, line items, currency, and counterparty. It then deduplicates against Xero and creates the correct bill or sales invoice via the Xero API, attaching the original PDF.

I designed and built the entire system: the LLM prompt and extraction logic, the Xero integration (dropping the official SDK for direct REST calls for reliability), a Swagger-documented API, and GitHub Actions CI/CD to Google Cloud Functions. Handles wildly different invoice formats with zero per-supplier parsing rules.

Kultura Research is a behavioral research platform where panel participants research a real purchase with an AI assistant, a Claude-based classifier tags their behavior, and researchers compare results across five markets. The client hired me through Toptal for a pre-launch security and code-quality review of the Next.js App Router, Supabase, and Anthropic API codebase. I delivered a findings report with 41 prioritized issues, each with a plain-English explanation, file and line references, and a ready-to-apply diff. Findings covered secret handling, Row Level Security design, an unauthenticated LLM proxy, GDPR erasure gaps, PII scrubbing on the wrong path, a confidential sponsor brief leaked in the client bundle, and race conditions corrupting transcripts.

I was then engaged to implement the fixes: HMAC-signed per-session tokens gating every LLM route, all participant writes moved server-side with a migration that makes the public key inert, server-side treatment/control assignment, a deterministic session closing flow, an abandonment beacon, ownership-gated deletion with an audit table, login rate limiting, an LLM-based journey-stage post-processor with a cron sweep, and a local integration test suite.