Senior Backend / AI Engineer — Architecture Review & Multi-Tenancy for a German ConTech Product

About us Jakob AI is an early-stage German company building AI tooling for architecture and engineering firms. Our product ingests project documents — structural reports, specialist input, expert opinions, IFC models — and generates construction tender documents (Leistungsverzeichnisse) from the firm's own historical archive. We have a paying pilot customer (a large German general planning firm) and two additional firms testing the product. Two founders, both technical, working part-time. We have a working system and real users. What we do not have is a second pair of experienced eyes on the foundation. The stack App: Next.js / TypeScript Backend: Python, FastAPI Retrieval: Dify (self-hosted, Docker), ChromaDB, Cohere embeddings Models: Anthropic Claude via AWS Bedrock (eu-central-1) Ingest: custom parsers for GAEB (X83/X84, D8x), IFC, Excel, PDF Infra: currently Hostinger/Hetzner, migration to AWS under consideration Two repositories: application and ingest pipeline What we need Phase 1 — Review (fixed scope, 3–5 days) A structured assessment of the whole system, delivered as a written findings report plus a prioritised backlog with effort estimates. Specifically: Architecture review. Is the current structure sound for the next 12 months and 10–20 customers? Where will it break first? Multi-tenancy. We currently deploy a separate instance per customer. We need a proper tenant model with server-side data isolation, since each customer's document archive is highly confidential and must never be reachable from another tenant. We want your recommendation on the model and a realistic migration path. Security. Authentication, authorisation, secrets handling, data at rest and in transit, dependency posture. GDPR matters to our customers, so data residency and processing boundaries are in scope. Code review. Not line-by-line. We want to know where the technical debt actually sits, what is fragile, and what should be rewritten versus left alone. AI pipeline. Our retrieval and generation path — chunking, embedding strategy, retrieval configuration, prompt structure, orchestration. Where are we losing quality for structural rather than domain reasons? Phase 2 — Implementation (optional, scoped after Phase 1) Depending on what Phase 1 finds, we would continue with the highest-priority items, most likely tenant isolation and the retrieval path. What we are not asking for Domain expertise in German construction. We have that in-house. A greenfield rewrite. We ship to real customers and cannot stop. Frontend or design work. Requirements 7+ years building and operating production backend systems Demonstrable experience designing multi-tenant SaaS with hard data isolation Strong Python; comfortable reading TypeScript AWS in production, ideally including Bedrock or a comparable managed model service Practical RAG experience — you have shipped a retrieval system and measured whether it actually works, not just built a demo Able to write a findings report a non-specialist stakeholder can act on Nice to have Docker, self-hosted infrastructure, migration to managed cloud GDPR-relevant architecture work for EU customers German language (helpful, not required — all documentation and communication can be in English) Experience with document parsing pipelines or structured data extraction How we work Paid trial first. Before the full review, we book one paid day on a narrow, defined piece of work. This works both ways: you see the codebase, we see how you work. Fixed scope and fixed price for Phase 1. No open-ended hourly billing. NDA before repository access. All work product assigned to us in writing. We will provide an architecture overview, data model, and our own measurement results up front. To apply Generic proposals will not be read. Please include: Start your message with the word "Gaggenau" so we know you read this. One multi-tenant system you designed: what isolation model you chose, and what you would do differently today. How you would evaluate whether a RAG system's problem is retrieval or generation — concretely, what you would measure first. A rough fixed price for Phase 1 and your availability. We read every application that follows these four points and reply to all of them, including the ones we decline.