The pilot never reaches production
The demo works, but quality, permissions, cost or operations are still unresolved.
01 / AI DEVELOPER & LLM FREELANCER BERLIN
AI and LLM products that work reliably in day-to-day operations.
I build AI functions for real specialist processes, from data and evaluation pipelines through RAG to controlled agents with tool access. Quality, permissions, cost and operations become measurable before production use.
THE PROJECT QUESTION
„How do we put AI into production without losing control of quality, cost or confidential data?“
MY ANSWER
Start with a tightly scoped use case, representative test data and an architecture that controls the model, data access and tool permissions separately. Measured results determine whether the system should be expanded.
WHEN I SUPPORT YOUR TEAM
When an AI pilot must deliver measurable value and survive production.
The first decision is not which model to buy. We clarify the task, data, acceptable failure modes and success criteria. That makes it clear whether RAG, an agent, conventional automation or a combination is commercially useful.
Internal knowledge remains difficult to use
Documents and specialist data need to become reliably accessible through RAG, with sources and clear access rules.
Sensitive data limits the choice of model
Legal, technical or internal data cannot be sent to external model providers without control.
The team lacks AI product experience
Your developers know the product; the missing experience is retrieval, agents, evals and local operation.
AI ENGINEERING / DECISION SYSTEM
From training data to inference: every layer is evaluated.
A dependable AI system does not emerge from selecting a famous model. Data quality, retrieval, model behaviour, quantisation and operations are designed separately and compared on representative tasks. Investment goes into the component that demonstrably needs improvement.
-
01→ Define the target, evaluation set and baseline first
Professional success criteria, difficult cases and acceptable failure modes come before architecture. A simple baseline shows whether RAG, an agent, fine-tuning or conventional software creates measurable value.
-
02→ Training-data pipelines instead of unchecked collections
I build pipelines for collection, cleaning, deduplication, normalisation, annotation, versioning and quality assurance. Training, validation and test data remain separate so overfitting cannot masquerade as progress.
-
03→ RAG, embedding models and vector databases as one system
Document processing, chunking, metadata, permission filters, embedding model, vector search and reranking are evaluated separately. This reveals whether retrieval or answer generation is the actual source of error.
-
04→ The smallest capable model wins
Frontier, local and specialised smaller models are compared on the same tasks. Qwen3.6-27B and Gemma 4 31B are possible local candidates, but latency, quality, privacy and operations decide, not a public benchmark.
-
05→ Quantisation is tested against quality and throughput
FP16, BF16 and quantised variants can change memory use, context, concurrency and answer quality. I measure those effects on the available hardware instead of treating quantisation as a free optimisation.
-
06✓ Fine-tune only when data and failure mode justify it
Fine-tuning does not repair poor or contradictory training data. Only after prompting, RAG and tool logic have been evaluated and a stable behavioural gap remains do I test SFT or adapter tuning against an independent regression suite.
DECISION RULE
No model choice without a task-specific evaluation. No fine-tuning without dependable data.
Quality, privacy, latency, throughput and operating cost are measured together. The architecture remains interchangeable when better models become available.
SELECTED PROJECTS / 2
Selected projects
in detail.
Real work, explained clearly. Confidential interfaces and data remain consistently anonymised.
CONFIDENTIAL / ANONYMISED
CASE-STUDYLegal Intelligence Platform
PROJECT VIEWRAGdomain-specific
DATAprivately processed
QUALITYreviewed by lawyers
Domain-specific RAG & rerankingACTIVE
Answers linked to concrete sourcesVERIFIED
Reference tests from real legal casesACTIVE
Client and role modelVERIFIED
SCREEN / 01Anonymised UI reconstruction · no client or case data
Legal Intelligence Platform
Accelerating legal research and preparation with verifiable sources.
A platform for a legal-sector company makes extensive legal sources accessible, structures matters and prepares working drafts. Several lawyers supplied real test cases and evaluated professional quality.
- CHALLENGE
- A plausible formulation is not enough in legal work. Sources, currency, jurisdiction and the reasoning path must remain traceable for the responsible lawyer.
- SOLUTION
- Domain-specific RAG pipelines combine adapted embedding and reranking models with source grounding, role permissions, structured quality tests and professional approval.
- OUTCOME
- Research and preparatory work are brought into one workflow while legal evaluation and approval remain transparently with the lawyer.
- ✓Domain-specific RAG & reranking
- ✓Answers linked to concrete sources
- ✓Reference tests from real legal cases
- ✓Client and role model
- ✓Private model options
LaravelLLMsEmbeddingsVector searchQuality tests
CONFIDENTIAL / ANONYMISED
CASE-STUDYAutonomous Pentest Agent
PROJECT VIEWMODELSlocal option
TOOLSisolated
EVIDENCEtraceable
Local models & owned infrastructureACTIVE
Isolated tool executionVERIFIED
Traceable reportsACTIVE
Provider-independent architectureVERIFIED
SCREEN / 02Anonymised agent console · fully synthetic target data
Autonomous Pentest Agent
A pentest agent that is not tied to one model provider.
An autonomous system plans tests, uses isolated security tools, evaluates evidence and documents reproducible vulnerabilities. The architecture works with different model classes.
- CHALLENGE
- Information about client systems and vulnerabilities must not reach external model providers. At the same time, agents need controlled, repeatable and evidence-oriented behaviour.
- SOLUTION
- Provider-independent orchestration, local inference, isolated tool execution, explicit action budgets and quality tests using real authorised cases.
- OUTCOME
- Strong results in realistic scenarios using local models, restricted tool access and clear human control.
- ✓Local models & owned infrastructure
- ✓Isolated tool execution
- ✓Traceable reports
- ✓Provider-independent architecture
- ✓Human approvals
Local LLMsAI agentsPythonDockerBurp Suite
AI-ASSISTED DEVELOPMENT
Deliver faster.
Control data risk deliberately.
AI can accelerate analysis, implementation, tests and documentation. Model choice follows the task, data class and required level of control.
01FAST
Frontier models for appropriate tasks
For non-sensitive data, powerful models accelerate research, implementation, tests and documentation. Results are reviewed like every other contribution.
- More iterations in the same timeframe
- Faster prototypes and migrations
- Manual reviews and automated tests
02CONFIDENTIAL
Local models for sensitive work
Models can run locally or inside your controlled infrastructure for confidential code, legal data or internal documents.
- No sensitive data sent to external model providers
- Provider-independent architecture
- Operations aligned with your privacy requirements
CLIENT FEEDBACK / AI DEVELOPMENT
What clients value
about the collaboration.
“Damian treats the client’s problem as a challenge and the result is always outstanding. He goes deeply into every problem until he finds a solution. Giving up is not an option for Damian.
“Damian is a real stroke of luck. Clear, open communication and completely transparent work make the collaboration a success. Our requirements were implemented competently and flexibly.
FREQUENTLY ASKED QUESTIONS / AI DEVELOPER & LLM FREELANCER BERLIN
What team leads want to know before a project.
Clear answers about the start, collaboration, data protection and budget before you invest time in a long sales process.
Ask your question directly ↗Do you work with existing development and specialist teams?+
Yes. I can own an AI component, lead architecture and evaluation, or support your team during implementation. Decisions and interfaces are documented so knowledge remains inside the company.
Does our data have to go to OpenAI, Anthropic or another external provider?+
No. Depending on the required level of protection, we can use local models, self-hosted inference or clearly separated hybrid workflows. External models are used only where the data class and agreed rules allow it.
How do you determine whether a RAG system or agent actually works?+
Representative tasks and success criteria are defined before implementation. We then measure source grounding, task success, errors, cost and policy violations with repeatable tests.
Can a limited budget be enough for a useful start?+
Yes, if the first use case is narrow. A focused pilot answers the riskiest questions first and avoids spending money on a technically impressive but professionally useless demo.
Why can experienced AI development still be cost-effective?+
Frontier models accelerate analysis, implementation, tests and documentation. That speed is combined with manual reviews and automated tests. Local models are available for sensitive work.
What does an AI or LLM project cost?+
Data quality, integrations and quality requirements matter more than the model name. After a short assessment, you receive a sensible first scope with assumptions and a budget range. A focused feasibility review or pilot limits risk before larger delivery.
Which local models and systems are available?+
A workstation with 256 GB RAM and two RTX 4090 GPUs is available for compact and medium-sized models. Two DGX Spark systems extend local memory for larger models and contexts. Candidates such as Qwen3.6-27B and Gemma 4 31B are compared at suitable quantisation levels using your tasks. Quality, latency, throughput and privacy decide, not a public leaderboard.
When does fine-tuning make sense?+
When a narrow, recurring behavioural problem cannot be solved reliably through prompting, RAG or tool logic and enough high-quality training data exists. Without clean data, a separate test set and a baseline, fine-tuning can preserve errors instead of improving quality.
How long does a robust first pilot take?+
A tightly scoped pilot can often be evaluated within a few weeks. The decisive factors are test data, integrations and the quality threshold. The expected result, stop criteria and next decision point are agreed before work starts.
THREE DISCIPLINES, ONE PARTNER
Does your project need more than
one discipline?
Ambitious projects benefit when modern AI, scalable product development and offensive security are considered together.