Practice 01 · Techology, Inc.
AI should actually do the work.
The position001
The useful version of an agent has context, permissions, tools, evidence, and a clear stopping condition. I build the surfaces and control systems that let AI move from conversation to responsible execution.
Representative systemsBuilt + designed
Analytics agents that inspect sources, write queries, build dashboards, and verify the rendered result across web, CLI, and MCP. The same organization scope and secure persistence path carry across every surface.
React · TypeScript · MCP · CLI · SQL · Render verificationSemantic profiling and correction flows with a fast default model, evidence-based escalation, hard spend limits, one bounded retry, and an explicit human-review terminal state when the system cannot resolve the answer safely.
LLM routing · Evaluation · Cost controls · Dagster · Human reviewAgent and embedded-product architecture that keeps the human actor, credential, host client, and access surface separate. Signed assertions, idempotent events, confidence labels, and careful telemetry make that separation usable.
OAuth · Signed assertions · Event identity · Embedded analytics · PrivacyOperating standardsNon-negotiable
Inspection and provenance travel with the decision instead of disappearing after a tool call.
Writes are bounded, canary-sized, and independently read back before scope expands.
The system can stop honestly when confidence, authority, or evidence is insufficient.
Continue exploringPractice 02
Data foundations↗