AI, data, and engineering

Perspectives on AI systems, data platforms, software engineering, and the decisions that matter in production.

Security, Data Compliance, and AI Governance

22 May 2026

Security, Data Compliance, and AI Governance

The cost of retrofitting security and governance into a live system is 6 to 10 times higher than designing it in from the start. Here is what governance by design looks like in practice across enterprise AI, data, and software platforms.

Your RAG Pipeline Is Throwing Away the Evidence: Multi-Modal Retrieval for Scientific Documents

8 May 2026

Your RAG Pipeline Is Throwing Away the Evidence: Multi-Modal Retrieval for Scientific Documents

Text-only RAG discards the charts, figures, and tables that carry the real evidence in scientific docs. Here is how we fixed it with vision extraction, structure-aware chunking, and figure-level citations.

Redact Before the Model: How We Took Insurance Support QA From 5% to 30% Coverage With Zero PII Exposure

24 Apr 2026

Redact Before the Model: How We Took Insurance Support QA From 5% to 30% Coverage With Zero PII Exposure

Most insurance support QA samples 5% of tickets by hand. We pushed coverage to 30% with redaction in-pipeline and multi-agent scoring against the actual policy.

Your Maintenance Spreadsheets Are Hiding Eight Figures of P&L

10 Apr 2026

Your Maintenance Spreadsheets Are Hiding Eight Figures of P&L

Maintenance run from spreadsheets buries your biggest controllable cost. Model it as a lakehouse data product, rank bad-actor assets, and make repair-vs-replace arithmetic.

Your Fault Detection Model Is Accurate and Nobody Is Listening

27 Mar 2026

Your Fault Detection Model Is Accurate and Nobody Is Listening

Fault detection in building systems dies from false-alarm fatigue, not from low accuracy. The model that wins is the one operators still trust on month six.

Off-the-Shelf OCR Always Wins the Demo and Loses in Production

13 Mar 2026

Off-the-Shelf OCR Always Wins the Demo and Loses in Production

Generic OCR reads clean invoices and breaks on real documents. The value is the pipeline: format-aware parsing, validation, confidence thresholds, and human review on failed fields.

Your Pricing Engine's Real Defect Is the Missing Audit Trail, Not the UI

27 Feb 2026

Your Pricing Engine's Real Defect Is the Missing Audit Trail, Not the UI

In commodity fuel distribution, an UPDATE that overwrites the old price is a silent compliance and revenue defect. The fix is an append-only price history beside the ERP, not a better screen.

vLLM vs Hosted APIs: The Break-Even Almost Everyone Gets Wrong

13 Feb 2026

vLLM vs Hosted APIs: The Break-Even Almost Everyone Gets Wrong

Self-hosting open LLMs with vLLM beats hosted APIs only past a throughput break-even. The cost lever is utilization, not list price. The math, when not to, and a hybrid router.

Single-agent RAG hits a ceiling. Multi-agent breaks it, then breaks you differently.

30 Jan 2026

Single-agent RAG hits a ceiling. Multi-agent breaks it, then breaks you differently.

Single-agent RAG stalls on multi-step, stateful, tool-using work. Hierarchical multi-agent clears the ceiling but adds latency, drift, and nondeterminism. Here is how we ship it anyway.

AI Governance Is a Guardrail Stack, Not a Policy PDF

16 Jan 2026

AI Governance Is a Guardrail Stack, Not a Policy PDF

Governance is a stack of enforced controls, not a document. Map each to a tool and the risk it removes, build by blast radius, and cut anything that cannot block a deploy.

Want us to cover a topic?

Tell us what you are working on. We write about the decisions and patterns that matter in production.

Book a call