What we adopt, what we avoid.

Pair and mob programming, retrospectives, paired reviews, small and frequent releases. Still the most effective frame for shipping software that lasts — and a multiplier once AI joins the pair.

In production · adtech, telecoms, retail

Write the expectation before the implementation. With an AI assistant, tests become the contract that bounds what it generates: the faster code is produced, the more that contract matters.

The AI proposes, critiques and refactors; the engineer decides and stays accountable. A tireless pair, not a subcontractor: review, tests and judgement stay human.

Design data flows as events from day one and keep batch for recomputation. Data arrives fresh, and a single model serves both real time and reporting.

In production · adtech, rail

Model the catalogue on established ontologies (the W3C's SOSA/SSN for sensors and observations) rather than home-grown conventions. Metadata becomes queryable, alignable and durable.

In production · defence

Agents that read schemas and SQL, propose quality rules and tune them from observed anomalies. With typed, validated outputs, humans arbitrate instead of hand-writing every check.

In production · online betting

Derive schemas (Protobuf, Parquet) straight from the C, C++ or Python code that produces the data, through analysis and specialisation, instead of maintaining them by hand. Powerful on heterogeneous sources, still demanding to tool.

In production · defence

Measure test quality by injecting mutations into the code. As more code gets generated, it's a good guard against complacent tests — it still needs to get fast enough for CI.

Closing the loop between decisions and measurements in production (pricing, allocation, prioritisation). The potential is real, but observability, policy safety and exploration cost must be under control first.

Accepting generated code without tests or review just moves the debt to later, faster. Alternative: test-first and human-AI pairing, where tests bound what the AI produces.

Infrastructure as code in a real language (TypeScript, Python): types, functions, tests. Cloudflare, AWS and Koyeb declared alongside the rest of the code.

In production · short-term rentals, SaaS

Durable workflows written as ordinary code: retries, timeouts and compensation without home-grown scaffolding. The right tool as soon as a process outlives a request.

In production · short-term rentals

An agentic coding assistant in the terminal: it reads the repository, runs the tests and proposes complete changes. Our everyday AI pair, always framed by tests and review.

Programmable typesetting, compiled in milliseconds and embeddable as a library. Replaces LaTeX for generating PDF, SVG and HTML from typed data.

The standard protocol for connecting AI agents to tools and data. One MCP server makes a system usable from Claude, Cursor and others without bespoke integrations.

Open-source, extensible, API-driven metadata catalogue. Solid once a real domain vocabulary is modelled in it; operating it still takes investment.

In production · defence

A Hadoop-era job scheduler, barely maintained, with no notion of data assets or fine-grained retries. Alternative: Temporal for application workflows, a modern orchestrator for data pipelines.

In production · retail

Edge compute with no servers to manage, and S3-compatible object storage with no egress fees. An API moved from Rails to Rust on Workers can run at near-zero hosting cost.

In production · environment

The Kafka protocol without Kafka's complexity: a single C++ binary, no JVM or external coordination, lower latency. Kafka clients, connectors and tooling work unchanged.

The reference serverless analytical warehouse: no infrastructure to size, standard SQL, and pay-per-use pricing that rewards good modelling.

In production · retail, luxury, adtech

The common foundation for running heterogeneous services and workers, with Kustomize to keep manifests readable. Keep it for workloads that don't fit serverless.

In production · retail, SaaS, adtech

Lightweight, fast, easy-to-operate messaging with persistence through JetStream. Excellent for IoT and service-to-service communication; weigh it against Redpanda once large histories must be replayed.

In production · rail

Access to frontier models inside the client's AWS perimeter: data stays in the account, existing IAM and billing apply. The right entry point when compliance comes first.

In production · online betting, short-term rentals

SQLite and a vector database at the edge, right next to Workers. Promising for low-cost AI applications; size and query limits still to be proven.

The Kafka protocol is still the streaming standard, but its reference implementation isn't: the JVM, coordination and running costs no longer pay off on new projects. Alternative: Redpanda, API-compatible.

In production · rail, online betting, retail

Powerful but expensive, and each building block (Lakeflow, Unity Catalog) ties data and processing a little more to the platform. Alternative: open formats (Parquet), an independent catalogue and portable processing with Polars and SqlGlot.

In production · rail

Cloudera, Hive, Impala, Sqoop: a heavy stack to run, evolve and staff. Alternative: object storage, open formats and serverless or single-node engines.

In production · adtech

Performance, memory safety and a type system that makes refactoring calm. Our default for services, the edge and demanding data tooling.

In production · environment, defence

The lingua franca of data and AI. With strict typing and validated models it holds up in production; we switch to Rust where performance matters.

In production · online betting, defence

End-to-end strict typing for interfaces and the edge. This site is the demo: Astro, React islands and Workers, all in TypeScript.

In production · short-term rentals

Validated data models, and AI agents with typed inputs and outputs. The most reliable way to plug an LLM into production code.

In production · online betting

Rust-powered DataFrames with lazy, parallel execution: most "big data" volumes fit on a single machine. A fraction of a cluster's cost, and testable like ordinary code.

SQL parsing, optimisation and transpilation across dialects. Essential for auditing queries, extracting lineage or moving between engines without manual rewrites.

In production · online betting

An excellent language whose ecosystem no longer keeps up: a costly 2 → 3 migration, flagship frameworks losing momentum, a shrinking talent pool. Alternative: Rust for strong typing and performance, Python for data and AI; Scala remains relevant for existing systems and teaching.

In production · retail, education

Most real-world volumes no longer need a distributed cluster, nor its operational and testing overhead. Alternative: Polars and SqlGlot over Parquet; Spark remains justified at very large scale.

In production · online betting, adtech

Moved to a commercial licence (BSL) in 2022: a legal and budget risk for any new project. Alternative: Apache Pekko for existing systems, simpler models (Temporal, async Rust) for new work.

In production · adtech, SaaS