Machine Learning · 3 min read
Data drift in production: knowing when a model has aged
Models don't fail with errors: they lose accuracy silently as data changes. How to monitor input and prediction drift with PSI, and decide when to retrain.
Engineering notes.
Technical explanations of the problems we meet while building complex systems. No shortcuts, with code where it helps.
- Machine Learning · 3 min read
Data leakage in predictive models: when a model looks too good
Data leakage makes a model shine in validation and fail in production. The most common kinds, how to detect them, and how to validate on time-ordered data.
- Fintech · 3 min read
Decimal arithmetic in financial systems: cents that never vanish
Binary floating point can't represent 0.10 exactly. How to represent money, choose rounding rules and split amounts without losing or creating cents.
- Applied AI · 4 min read
Evaluate retrieval before generation: measuring a RAG system
Most RAG errors start in retrieval, not in the model. How to build an evaluation set and measure recall, MRR and faithfulness in CI.
- Blockchain Engineering · 3 min read
Handling chain reorganizations in EVM indexers
Indexers that treat recent blocks as final accumulate orphaned data. How to detect reorganizations, roll back safely and expose confirmation levels to consumers.
- Backend Engineering · 3 min read
Idempotency keys in APIs: preventing duplicate charges and orders
Retries are unavoidable on real networks. How to implement idempotency keys in an API: storage, concurrency, payload validation and expiry.
- Fintech · 4 min read
Pre-trade risk checks: designing the gate on the order path
How to design risk controls that run before every order: which checks to include, in what order, how to keep state without locks, and why to fail closed.
- Applied AI · 3 min read
Safe AI agents: tools, permissions and human approval
An AI agent is only as safe as the tools it can call. How to scope tools, propagate user permissions and require approval before real-world effects.
- Blockchain Engineering · 3 min read
Smart contract upgrades: proxies, risks and when not to use them
Upgradeable contracts solve one problem and create others. How proxies work, the most common storage mistakes and the governance an upgrade requires.
- Software Architecture · 3 min read
The modular monolith before microservices
Microservices solve scaling and organizational problems, and create network and consistency ones. How to build a modular monolith that can be split later.
- Software Architecture · 3 min read
The transactional outbox: publishing events without dual writes
Writing to a database and publishing to a broker fail independently. How the outbox pattern fixes this, with a relay, ordering and idempotent consumers.
- Backend Engineering · 3 min read
Zero-downtime migrations in PostgreSQL
A seemingly simple migration can lock an entire table in production. How to avoid long locks, use expand/contract, and add indexes and constraints safely.