Gösta (Von) Stebut-Freelance Software Developer & Architect

• Own product initiative — AI Workspace: a private AI research and knowledge platform for European professionals who work with confidential material — lawyers, consultants, compliance leads, founders. The premise: confidential documents should become usable for AI without leaving the customer's control, and without a US-controlled company in the data path (CLOUD Act / FISA 702).
• Autonomous deep research engine: A multi-step research pipeline that breaks a question into sub-questions, runs them in parallel across web and private sources, deduplicates and cross-checks the results, and produces a fully referenced report. Built in Go with an adversarial verification layer: every result is independently challenged before it may enter the report — that is exactly what separates a usable research result from something that just sounds plausible.
• Private knowledge library with client-side encryption: Documents are encrypted on the user's device (libsodium) and remain readable for no one else — not even the operator. The real engineering challenge is making encrypted content searchable and usable for AI at all: solved through a hybrid retrieval layer (pgvector embeddings plus lexical search) on a key model where the key stays with the client.
• Agent orchestration and workflow automation: A workflow engine that chains agent steps, tool calls, and external connectors. For this, I designed and implemented a send-gate security layer: every action is classified as read / write / send, and any step that would send data to third parties pauses until a human explicitly approves it — with anything unclassified falling back to the safest interpretation. That is the architectural answer to the main objection against autonomous agents in a professional setting.
• Full platform and operations: Go services (Fiber, gRPC, Protobuf) with Next.js / React / TypeScript frontend, PostgreSQL + pgvector, Redis, BadgerDB, S3-compatible object storage. Three environments (dev / preprod / prod) on self-hosted EU infrastructure, deployed via Docker Compose and Traefik, with Prometheus / Grafana / Loki / Jaeger for observability and full GDPR documentation. Around 20,000 commits, built and operated by me alone.

• Prototyping the SIMDA core logic (sense-infer-mind-decide-act): Building the first proof of concept that captured customer responses and derived knowledge from them in the form of a numerical representation of a customer's "mindgraph" attributes — emotions, styles, preferences. The challenge was a small codebase with low latency that the founder could experiment with at the same time: I implemented the backend in Go (robustness and speed), intentionally kept the frontend to plain HTML/CSS/JavaScript, and turned the founder's YAML configuration into code plus task-specific configuration so changes could go through proper CI/CD and change management. A fashion e-commerce customer used the prototype against its live customer base for next-best-offer recommendations.
• Big-data cloud architecture for a cosmetics group (Azure): Two years of consulting for a large makeup group that needed a cloud architecture for its big-data stack. I advised on the selection of cloud services, implemented CI/CD for the Azure Kubernetes offering, and built a complete operating setup — metrics, logging, dashboards, alerting. For managing multiple clusters (website and related backend services), I implemented GitOps and Infrastructure as Code based on GitHub, CUE, and Argo CD, designed for speed and reliability with automatic failover for local and regional Azure outages.
• Leading 5-10 developers for low-latency personalization backends: WUNDER built a low-latency backend that personalized a large e-commerce website using SIMDA logic. I led a mixed team of employees and external developers that implemented a scalable Go microservice system on Kubernetes for first- and second-touch page personalization. Hard requirement: decisions in under 5 ms under production load — with a team that was learning Go in parallel. I ran code reviews, group and one-on-one training, and enforced test coverage; that was the mechanism that made this combination workable.
• Leading 4 developers for the Mindgraph data pipelines: WUNDER's internal knowledge data was processed through Python pipelines. I provided the team with guidelines, Dockerfiles, and parameters to break the pipelines into manageable building blocks, and built the workflow definitions (YAML, CUE) for Argo Workflows around those blocks. This produced workflow patterns that served as self-service templates for the data scientists — afterwards they could create additional pipelines without me. I was also responsible for DevOps and architectural support for this system.
• Leading 4 developers for WUNDER Playtime: A public, multi-tenant product for playful consumer interaction with the Mindgraph. Under my guidance, a highly customizable React frontend was created; the existing SIMDA system was reused and adapted as the backend. To optimize the frontend, I used behavioral tracking via Matomo — the biggest challenge was that the application often ran inside embedded custom browsers (TikTok and similar), where the usual assumptions about the environment do not apply.
• Leading 4 developers for WUNDER Strategymind: An analytics product that lets customers run evaluations on their brand, their products, and their customer base. Under my guidance, a React frontend was created; the existing backend systems (data pipelines, workflows) were combined for job processing. My specialization here was prompt engineering for integrating OpenAI GPT-3 and GPT-4 into these pipelines, implemented through Go applications against the Azure OpenAI Service.

• sendandstore.de: Liquid Labs develops startup ideas for the OTTO Group; I was hired for the technical leadership of a B2C logistics business model. I was directly involved in shaping the business case, the service offering, and the operational processes. I led a team of 2-5 developers in implementing the services and provided DevOps support for the deployed systems. The hardest part was building a system while the business idea kept changing — and at the same time coordinating integration with OTTO Group services (Hermes Logistics, warehousing).

• Expansion of the dispatch control system FELIS: FELIS is a dispatch control system used by several German police authorities based on C/C++. I expanded it in the frontend and backend and implemented many customer change requests.
• Design and implementation of TETRA support: New requirements around BOS digital radio TETRA forced design changes: many services were rewritten in C#, and the old CORBA RPC system was replaced with ZeroMQ and Protobuf. I designed and implemented a large part of the backend systems for the new architecture.
• In this context, I received the VS-NFD authorization (classified information - for official use only) for police documents — expired in 2023.

Consulting and development of websites, connection to internal systems, network infrastructure, VPN

Lecture series on Business Intelligence - MS SQL Server 2005, Integration / Analysis / Reporting Services, Data Mining

Solution Designer for Lawful Interception, network design, Cisco, Utimaco, VPN, IPSec, accounting

VoIP phone systems

Savings bank IT, Java, Basel II credit risk assessment

Consulting on the design of a Basel II credit risk assessment