AI-Wall · Michał Mazurowski, Solutions Architect · Gdańsk, Poland — remote
Dozens of external sources, each with its own format, its own gaps and its own idea of what a date is. I bring them into one model you can query, put an API over it, and take responsibility for it still working tomorrow — with monitoring, releases, and a procedure for the day a source disappears or changes its schema without telling anyone.
These are not three services to choose from. They are one path: data comes in, becomes a service, and the service keeps working. Most clients have the first step behind them and are stuck on the third.
External sources do not agree with each other, and they never will. They differ in format, in units, in coordinate system, in what the same field name means, and in what they do when they simply do not know. I build the ingestion layer that brings them into a shared model without discarding the original attributes — because “where did this value come from” always gets asked, just later.
PostgreSQL / PostGIS · FastAPI + asyncpg · materialised views · partitioning · JSONB
Evidence: scientific and governmental registries, each with its own format and its own gaps, brought into one queryable model with source attributes preserved. Demonstrated on environmental and geospatial data — that is the domain I built it in, not the limit of where I can take it.
A release has to be repeatable and auditable: an approval gate where the risk warrants one, and no human where it does not. The same rule applies to AI. I do not sell “agents” — I build them, but what I sell is a working system, not the technology. I put models into production under a regime: guardrails, an audit trail, measured cost, and an answer to what the system does when a source is missing.
GitHub Actions · GitLab CI · Jenkins · Azure DevOps · Terraform · Kubernetes · Docker · AWS · Azure · Cloud Run
Evidence: a controlled benchmark of five model configurations on one identical task, each in its own worktree off the same commit, with a byte-identical prompt — measuring cost, number of turns and passing tests. Which model and which settings is a question I answer by measurement, not by instinct — and the largest single difference came from one parameter rather than from switching models. Separately: an export gate with an allow-list of paths, tested by deliberate self-sabotage and run in CI.
This is the part without which the first two steps are just a claim. Logs, metrics and events in one place, so a failure is visible to you before it is visible to your customers. Plus the unglamorous half that actually decides outcomes: triage, prioritisation and reporting against an agreed service level — and telling apart what stops the fire from what removes it.
ELK · Splunk · Graylog · Solr · production monitoring · SLA · root-cause analysis · runbooks
Evidence: over twenty years on the operations side — from a service dispatch desk and night shifts through to architecture. “I will keep running what I built” is only credible from someone who has spent time on that end of the phone.
Five things we settle before the first line of code.
AI-Wall is one person today, and I say so plainly, because it changes how the work goes: you talk to the same person who then writes the code and answers the phone when something stops.
Over twenty years of production operations — from the service dispatch desk and network operations, through Linux systems engineering and deployments under audit, to cloud and AI solution architecture. At one employer I moved through four successive roles in five and a half years; twice I led a team, though leading was never the main job — I would rather answer for a system that runs than for a rota. I work alone, as a subcontractor, within a narrowly defined scope. Registered as an independent researcher: ORCID 0009-0007-3786-0310.
Selected work from my employment history, plus an independent research project that is public and can be checked without taking my word for it. I do not name employers or clients here — I will discuss them in a conversation, as far as the agreements allow.
Non-commercial project
The ground where the whole claim above was demonstrated end to end. A research platform on deep-sea and land mining transparency, built and run by me single-handed: ingestion from dozens of mismatched registries, API services, database, frontend, containers, releases and monitoring. It is not a product and not a service, and access to it is not for sale — it is evidence. The methodology is written up in a citable methods paper with a DOI.
The easiest first step: tell me what keeps breaking — one source that wrecks the import every week, or a deployment nobody wants to touch. I will tell you whether it is something for me, and what I need in order to quote it properly. You do not need a finished specification or a decision to work together. I share a phone number in reply to your message.