Research notebook / Connected studies
Follow the question.
How do we build AI systems whose data, decisions, and behavior remain understandable under real operational constraints?
Contemporary studies revisiting foundations from June 2024 through September 2026. Read across the series, or start with the constraint closest to your work.
Archive dates organize this retrospective collection from June 2024 onward. Each article separately records when it was added online and last revised.
- 18
- published studies
- 6
- research series
- 2024–26
- foundations revisited
Data systems and migration
From JDBC to AI-ready data: migrating without breaking existing consumers
A migration design built around compatibility contracts, change capture, reconciliation, and a cutover that can actually be reversed.
The database boundary in an AI system
Follow one entity through Postgres, Redshift, Cassandra, and a vector index to make ownership, freshness, and semantic meaning explicit.
Before adding compute, inspect the Parquet layout
A measurement plan connecting small files, row groups, partition choices, and compaction to the real work performed by an analytical query.
When a replay becomes a duplicate: Postgres, Kafka, and the commit boundary
A failure-driven design for transactional outboxes, consumer receipts, and recovery that preserves business effects when messages arrive again.
Dependable systems
AegisOS: what an AI system can promise when the network disappears
Local evidence, signed timelines, delayed transport, and the difference between a working offline prototype and a validated field system.
Python, Rust, and Go at the boundaries of an agent system
A design study of deadlines, backpressure, tool contracts, and ambiguous failures before deciding whether a language boundary is worth adding.
Evaluating change impact when the dependency graph is incomplete
A general evaluation protocol for separating dependency coverage, impact ranking, and verification, with explicit limits on what each can establish.
Governed AI and evidence
Structura: preserving evidence before documents become embeddings
A source-to-answer design study of document identity, extraction, retrieval, and citations—and the information an embedding cannot preserve.
When retrieval should retry, abstain, or ask for review
A corrective retrieval protocol that treats missing evidence, contradiction, permissions, and exhausted budgets as different decisions.
Semantic caching when permissions and documents keep changing
Why similar questions are not necessarily equivalent requests, and how evidence lineage changes cache lookup, invalidation, and evaluation.
Governed Bedrock applications: tracing a request from identity to evidence
A request-level design for retrieval permissions, model access, tool authorization, network boundaries, and audit evidence in AWS.
From biomedical text to research evidence
A methods note on biomedical concept extraction, ontology, provenance, and why matching a medical term is not the same as validating a claim.
Learning and evaluation
When does a neural representation earn its complexity?
A study protocol comparing simple baselines, frozen embeddings, and learned representations under the same data, shift, and deployment constraints.
A better fraud score can hide a worse model
An experiment protocol connecting temporal splits, SMOTE, probability calibration, and the workload an investigator can actually handle.
One ML pipeline, three execution layers: PySpark, Polars, and Rust
A reproducible experiment for choosing distributed preparation, local iteration, and native processing around a SageMaker training workflow.
Turning engineering questions into reproducible studies
A research agenda built around bounded claims, meaningful baselines, traceable experiments, and results that remain useful when a hypothesis fails.