Software architecture that holds.
Software architecture and engineering for backend, data and AI systems: eight years of keeping production in order.
Scroll to build
I. Built from the backend up.
Since 2018: freelance work, QA automation, then full-stack engineering in product teams. Today PaxCode contracts B2B: Node.js and TypeScript services, AWS, and data pipelines in production.
- F / 01 Backend
- F / 02 Data
- F / 03 AWS
II. One drum per system in production.
Each project below carried real load, with decisions and trade-offs documented.
- Drum 03 LLM pipelines: from a transcript to a reviewed document For a UK AI fintech. The patterns behind LLM pipelines that turn transcripts into documents a person reviews. Read case study
- Drum 02 No-code template editor Proposed and built for the same client. Output defined as data, not code, so customers change it without a release. Read case study
- Drum 01 Data leak scanner Searching the open web for leaked personal data. Distributed across machines and threads, rebuilt and run solo. Read case study
III. AI that does the paperwork.
The current focus is AI in production: LLM pipelines over meeting transcripts, prompt design, multi-modal processing and caching that keeps each run fast and affordable. Plus the tools that let non-engineers build pipelines of their own.
- I. Architecture review
- II. Backend & data engineering
- III. AI & LLM pipelines
- IV. Product & UI/UX design
- V. Senior engineer on contract
- End to end, from the data model to the interface, one owner.
IV. From raw input to a finished answer.
One adviser meeting, followed through an example AI pipeline: twenty steps from a recording to a reviewed report in a CRM. Keep scrolling to run it.
- 01 Sources audio · pdf · api
- 02 Parse STT · OCR · diarise
- 03 Normalise schema · metadata
- 04 Validate dedup · checks
- 05 Event queue SQS · retries · DLQ
- 06 PII redaction NER · rules · GDPR
- 07 Chunking semantic · overlap
- 08 Embeddings batch · cache
- 09 Hybrid index vector · BM25
- 10 Knowledge base versions · filters
- 11 Agent orchestrator plan · act · observe
- 12 Retrieval (RAG) top-k · rerank
- 13 Context engineering tokens · memory
- 14 Tools MCP · functions
- 15 LLM gateway routing · cache
- 16 Guardrails input · output
- 17 Structured output schema · retry
- 18 Human review human in the loop
- 19 Delivery SSE · docs · API
- 20 Client systems CRM · webhooks
V. Order in complex systems.
Pax Romana meant two centuries of stability. The goal here is more practical: systems that stay understandable as code, data and AI pile up.
- 01 Decisions written down, trade-offs stated.
- 02 Numbers over adjectives.
- 03 The simplest design that survives the next order of magnitude.
- 04 Least privilege by default: every service, role and key gets only what it needs.
- 05 AI output is checked by a person before it becomes a decision.
Latest writing
All postsVI. Let’s look at your system.
A 30-minute call: what you’re building, where it hurts, and whether PaxCode is the right fit. Based in Poland, working remotely with teams in the UK and EU.