Local-first agents and evaluation-led delivery
Applied ML Systems
Personal / Side ProjectExtending platform engineering habits into applied ML workflows with privacy and reliability as first-class constraints.
- Prototype local agents that keep sensitive context on-device whenever practical.
- Design evaluation loops before scale-up so quality decisions remain explicit.
- Carry forward data-contract and observability patterns from production platform work.
Impact
Creates a practical path from experimentation to trustworthy ML-assisted product workflows.
Systems Built
Local orchestration patterns for task agents; Evaluation harness patterns for prompt and behavior regression; Contract-first data flows between app layers and model interfaces
Constraints
Privacy-first handling of user context; Deterministic fallback paths when model confidence is low; Clear operator visibility into failures
Tooling
TypeScript · Node.js · Prompt evaluation patterns · Telemetry instrumentation