Reference builds
Runnable, written-up seams around production agent systems. The point is not to show a demo; it is to make the architecture, evidence, and operating pattern concrete.
Start with the operating question
Use a problem path to diagnose the missing contract, or a pillar to see the broader architecture. The linked builds are bounded implementation evidence—not proof of deployment, scale, compliance, or client outcomes.
- Our AI demo works, but production is stuck — diagnose the production gap before choosing a build pattern.
- Our agent is too unreliable to operate — connect checkpoints, idempotent effects, cancellation, and replay to the durable-execution build.
- We cannot trust what the agent remembers — connect governed records, scope, temporal history, and forgetting to the memory build.
- Production Agent Architecture — place durable execution inside the harness around the model.
- Agent Memory as Data Architecture — place semantic retrieval inside a governed data lifecycle.
Published
-
Verified-capability routing for multi-agent systems
Dynamic agent selection based on verified capabilities and trust boundaries.
Agentic AI -
Agent sandbox sized to blast radius
Capability-based isolation with resource limits, egress controls, and audit trail for untrusted agent code.
Platform Engineering -
Enterprise RAG document extraction pipeline
High-volume PDF/scan ingestion with extraction coverage audit and human-in-the-loop exception handling.
Data Platform -
Durable agent execution with checkpoints and cancellation
Resumable runs, idempotent tool effects, explicit cancellation, and replayable failure evidence.
Agentic AI -
Agent memory over a lakehouse
Public implementation of typed, scoped, bi-temporal memory records; 59 tests and six scenarios rerun on 3 September 2026.
Agentic AI -
Evaluation and observability harness for agent systems
One typed run ledger projected into production trace, replay fixture, eval result, drift signal, and cost report.
Agentic AI -
Sutra Feed Guard
Governed change control for external CSV and JSON feeds, with deterministic dispositions and reproducible incident evidence.
Data Platform -
Agent privacy as a data-flow architecture
Three enforceable privacy boundaries — context, tools, and memory — sharing one reconstructable decision ledger.
Agentic AI -
Exception queue before happy-path automation
Structured human handoff, policy-driven escalation, recovery states, and an unresolved-work ledger.
Agentic AI -
Policy-as-code for enterprise agent governance
Versioned rules, approval gates, deny precedence, runtime enforcement, and audit evidence.
Platform Engineering -
LLM gateway as the control plane for agent systems
One policy snapshot composes model eligibility, quality, residency, spend, fallback, and routing evidence.
Platform Engineering
Current portfolio
- Eleven reference builds so far. This portfolio will grow as new production seams and use cases emerge.
Future work
- New case studies will be added as useful production use cases emerge.