Posts
accessibilityagentsaiapiauditauthorizationawsbigquerycareercedarclassificationclaude-codeclicompliancecontext-engineeringcursordatadata-governancedatabricksdbtdeveloper-toolsdevtoolsdistributed-systemsembeddingsengineeringenterpriseevalsgcpgdprgovernancehipaaiacimmutainferenceinterviewiso27001jsonkuberneteslake-formationlearning-notesllmonetrustopaopen-sourceorchestrationperformancepolicyprivaceraprivacyretrosecuritysnowflakesoc2softwarestreamingstructured-outputstructured-outputssystem-designterminalterraformthurtool-usetoolingtoolsunity-catalogverificationvision
2026
- Reading DeepSeek DSec: agents build the environments that train agents
- Decision models: accuracy converges, calibration does not
- Verified erasure: lineage as an executable erasure program
- Reading software factories: artifacts on the edges, humans at merge
- Long-horizon agent state: the dimensions, and what moved in 2026
- Procedural graphs: forcing the plan is worse than no plan
- Managing agent skills in an engineering org: personal outranks project
- Terminals became agent consoles. Only one shipped a policy gate.
- Open weights dissolve ZDR. They do not replace EFS
- Which Cursor model should I use, and what each one is for
- Splitting custody from detection: EFS, Private Safety Processing, and ZDR
- There is no golden-standard policy language
- Reading Anthropic's system prompts: the date moved to the end
- Reading the harness debate: the prompt shrinks, the gates don't
- Hallucinate the tag, then repair it: measuring the baselines
- Reading Anthropic on multiagent systems: coordination is not emergent
- Reading ollama launch: one verb, nineteen bespoke adapters
- Reading pi's agent-loop.ts: the hard parts are all bug-shaped
- The orchestrator is blind on purpose
- Sizing the guard model: why auto mode uses Sonnet, not a small classifier
- How I would design an LLM inference API in a system design interview
- Notes on inference architecture: the trade-offs don't transfer
- Governing multi-party agentic systems
- Reading kimi-k3-in-c: 2.8T parameters in 8GB of RAM
- Why coding agents would rather not look
- Why coding agents lean on the shell
- The schema is the contract: structured output from protobuf to agents
- Why some AI demos resolve text instead of typing it
- The verification loop decides how many agents you can run
- Skills vs. subagents: when to use each in coding agents
- How streaming works in LLM chat and agentic systems
- How agents handle structured I/O
- What RTK Does to Agent Shell Output
- Where the Auditor Will Find Gaps in Your Data Platform
- Data Governance as Code: dbt + Terraform Patterns
- Data Policy Overlay Vendors: Immuta, Privacera, OneTrust, and Lake Formation
- Data Protection in Databricks Unity Catalog: A Field Guide
- Data Protection in BigQuery: A Field Guide
- How Modern Data Platforms Protect Data
- AI Coding Interview Rubrics