Maximilian B.-Freelance AI Systems & Platform Architect

Multi-Agent Document-Generation Platform

Problem:

Turn a single prompt into a well-researched, structured and cited document, while keeping the user in control: every stage can be reviewed, corrected or revisited before the system moves on.

Architecture:

A multi-agent system orchestrated by a central controller. More than 20 specialised pydantic-ai subagents (understanding, structuring, planning, research, writing, assembly, cross-reference check, glossary, citation, reference table and more) pass results forward through a shared plan persisted in Postgres (asyncpg); a runner layer dispatches each task by name to its agent, and mechanical steps such as formatting, table of contents and list of figures are deterministic code instead of LLM calls. Provider (Anthropic or OpenAI) is a configuration setting, and responses stream over Server-Sent Events.

Hosting & Infrastructure:

Stateless container run as a Scaleway Serverless Container on a private network with managed PostgreSQL, object storage for finished documents and assets, and centralized log shipping (Loki / Scaleway Cockpit), provisioned with OpenTofu modules and deployed through GitLab CI; usage is metered per API key.

My Contribution & Key Decisions:

• Central decision controller: a satisfaction classifier turns the user's free-text feedback into a verdict (keep working, advance, cancel or revise an earlier task), and a middleware applies it to the plan. Plan progression supports jumping back to earlier tasks and re-running the stages after them.
• Research and writing fan-out: planning produces a plan from which a research-design agent generates individual research tasks and a writing-design agent generates individual writing tasks, each linked to the research it needs. Every task runs in its own dedicated agent with its own conversation history.
• Self-review and human-in-the-loop: after each step a critical-thinking agent reviews the output; the user can accept, refine, or ask questions about a critique before the plan advances.
• MCP integrations: two remote MCP servers, web search and internal document search, connected through pydantic-ai with retries and long-running tool-call timeouts. Classifier agents decide per task whether web search is needed, and a small model condenses raw search results into structured findings before they reach the writing agents.

Usage Metering:

Clients authenticate with an API key validated against a key-management service; consumption is charged in credits calculated from token usage per model plus a markup, and exhausted quota is answered with HTTP 429.

Google Workspace MCP Server

Problem:

Give AI agents authenticated access to a user's Google Drive, Gmail, Calendar and Contacts without writing one tool per API, and make sure an agent can never change or delete data unless the user understands and approves the action.

Architecture:

A FastMCP server with three tool namespaces: a generic gateway (read from, and write to, any validated googleapis.com endpoint), a meta namespace (granted scopes, user profile) and a path namespace (Drive path index and migration mappings). Each user signs in with Google OAuth 2.0 and sessions are carried as signed JWTs; Google tokens are Fernet-encrypted at rest in Redis. Structured data lives in Postgres through async SQLAlchemy with Alembic migrations.

Hosting & Infrastructure:

Scaleway Serverless Container on a private network with managed Redis and PostgreSQL, provisioned per environment (dev, test, prod) with OpenTofu modules and deployed through a GitLab pipeline with Trivy, Ruff and ty gates; logs and OpenTelemetry traces ship to Scaleway Cockpit.

Key Decisions:

• Generic gateway instead of a tool per endpoint: two gateway tools cover Drive, Gmail, Calendar and Contacts, so the tool surface stays small. Input models validate that every request targets googleapis.com.
• Confirmation-gated writes: every create, update or delete is first described in one plain-language sentence by a small LLM (for example what will be permanently removed), and the user can accept, decline or request changes through MCP elicitation. The request is only executed against Google after an explicit accept.
• Safe fallback for clients without elicitation: a two-step flow returns the confirmation text and an event id while the request is stored per user in Redis; the second call must repeat an identical request, which is verified against the stored payload before it runs, so an approved action cannot be swapped for a different one.
• Drive path index and migration mappings: Postgres tables for drives, paths (with soft deletion) and old-to-new location mappings let an agent answer where a file lives or moved to after a company-wide Drive reorganization. Two notebooks populate them: one enumerates all Shared Drives through the Drive API (building full paths by walking the parent chain) and parses the mapping spreadsheet, the other validates references and upserts idempotently in batches.

Company-wide AI and data lead; led a 5-FTE team plus working students.

Led data science initiatives end-to-end, from data engineering and architecture design to deploying production AI solutions. Combined hands-on technical work with project management and professional as well as disciplinary team leadership.

• Designed and implemented a lakehouse architecture on Azure Data Lake Gen. 2, integrating data from SAP, Google Analytics, and Jira into a unified analytics platform
• Established a modern reporting landscape for international stakeholders using Power BI, eliminating most manual reporting within two years

Domain-Specific AI Translation Platform

Problem:

Replace expensive external translation of product data, technical documentation and web-content JSON for a global machinery manufacturer, in 100+ language combinations, with consistent domain terminology and approved wording.

Hosting & Infrastructure:

Services containerized with Docker and run on Kubernetes on Azure, exposed through a REST API with additional batch and queue-based processing for bulk jobs. SAP consumed the platform as a hosted service behind VPN; further connectors served on-premise systems and third-party cloud software. Model training for the second generation ran on on-premise GPUs.

Three Generations:

• Generation 1, custom-trained neural MT. Transformer/seq2seq models trained on existing translation memories, glossaries and technical documentation. Replaced because translation quality was not good enough and domain training caused catastrophic forgetting in general-language text.
• Generation 2, fine-tuned open pretrained MT models. Fine-tuned on on-premise GPUs. Replaced because every change to approved translations required retraining, the effort and cost per update were high, and stronger commercial LLMs had become available.
• Generation 3, retrieval-augmented generation on commercial LLM APIs. Approved translations are stored with embeddings in Postgres + pgvector. For each segment, an exact-match shortcut reuses already-approved translations without an LLM call; otherwise the most similar approved sentences are retrieved and passed to the model as few-shot examples. Updating approved wording now means updating data instead of retraining a model.

Key Decisions:

• Knowledge in data, not in weights: moving approved translations out of model parameters into a vector store made corrections immediate and removed the retraining cycle.
• Provider independence: OpenAI and Anthropic models in production with fallback between providers, so an API outage or model change does not stop translation.
• Structure-safe JSON translation: only values are translated; keys and structure stay untouched.

Quality Assurance:

Human review by language experts, terminology compliance checks, and acceptance feedback from the business departments using the output.

Outcome:

€2.7 million of external translation costs avoided over two years (words translated × agency price per word); 5 million+ translations served across product data, technical documentation and web content.

Microservice architecture, integrations and business-process automation for SMBs across industries.