System Design
The Kam Intelligence Registry


Kam AI
Product and research

System Design


Kam AI
Product and research

Kam needs one map that shows how everything connects.
The map starts with a workload ID.
It connects the answer trace, the label, the judge result, the fixture, the release gate, and the work item.
That map is the Kam Intelligence Registry.
The most important product object in Kam AI is not the model response.
It is the relationship between evidence objects.
A failed answer is only useful if Kam can connect it to the workload that produced it, the label that explains it, the fixture that protects against it, and the release gate that proves the fix stayed fixed.
That is why Kam needs an Intelligence Registry.
Without a registry, AI quality work becomes scattered.
One trace is in logs. One label is in a review queue. One fixture is in a test directory. One judge result is in an eval report. One release gate is in CI. One work item is in an ops board.
Each object may be valid, but the system cannot answer basic questions:
The registry turns those questions into normal product queries.
Visual artifact
The registry connects evidence by workload, then lets KamOps and KamSRE render the right view for each human workflow.
A production answer records route, hot reads, source context, tool plan, and answer metadata.
A reviewer states what should have happened: intent, entities, route, reads, source rules, and expected answer shape.
Approved evidence becomes regression coverage with deterministic graders and expected contracts.
Release gates evaluate workload health before risky changes move forward.
The registry should be compact. It should point to full artifacts instead of copying everything.
Registry object types
Takeaway: Keep indexes small and artifacts durable. The registry should help humans navigate evidence, not become a blob store.
The registry should not become a backend data-ops control plane.
For Kam, the cleaner split is:
That keeps the Next.js dashboard focused on viewing and approving bounded objects through API routes. It does not run cleanup, archive, scheduler, or import orchestration from the UI.
Workload ID is the stable unit of quality.
Routes change. Prompts change. Model providers change. UI screens change. But the product obligation remains.
chat.team_trends.v1 still means the user expects a team trend answer with the right sport, team, sample, dates, denominator, source context, and caveats.
agentic.trace_label_prep.v1 still means the internal workflow should read a failed trace, draft label expectations, run deterministic checks, and hand off a review item to a human.
What workload IDs unlock
Compare quality, latency, cost, and failures by product workload instead of raw endpoint.
Group repeated failures into one review packet instead of making humans inspect every duplicate.
Build eval sets from approved labels tied to the same product obligation.
Block changes that damage a specific workload even when global tests still pass.
Show whether one lane is getting healthier or drifting.
Create bounded internal tasks that know what workload they are improving.
Takeaway: A workload ID is the handle that turns AI behavior into product operations.
Trust receipt
A useful answer should leave a small receipt: route, scope, freshness, evidence, missing data, and confidence state.
Route
chat.team_trends.v1
Scope
A team trend answer that must expose the games, dates, spreads, scores, and ATS outcomes behind the aggregate.
Freshness
Trace and label evidence are current; release-gate enforcement is not active yet.
Evidence loaded
Missing or caveated
The registry is the backbone of the better Kam framework.
It makes trace evidence searchable, label work repeatable, fixture promotion visible, and release quality measurable. It also keeps open-source integrations in their right place. Langfuse, Phoenix, Label Studio, Argilla, promptfoo, or Evidently can help with slices of the workflow, but the registry remains Kam-native because the product taxonomy is Kam-native.
The next action is clear:
Build the registry v1 around workload ID, keep it as a derived projection, and make every evidence object link back to the product obligation it supports.
Read next
Why Kam is moving from AI chat into a production loop of traces, labels, graders, fixtures, release gates, and agentic work.
9 min read
Why Kam treats workload IDs as the unit for trace slicing, deduplication, dataset creation, and release scorecards.
7 min read
How Langfuse, Phoenix, OpenTelemetry, Label Studio, Argilla, promptfoo, DeepEval, Evidently, Temporal, Dagster, and LiteLLM fit around Kam.
9 min read