Systems that do the work, not just suggest it.03 / 06

Automation & AI

Agents with real tool access, retrieval you can audit, and an evaluation harness that runs before anything ships.

ic-agents · support-triage · eval run 218480 held-out cases · candidate-b · promotion blocked until green

evals/support_triage.py

1from ic.evals import Harness, Dataset, Scorer2from ic.agents import SupportTriage34ds = Dataset.from_jsonl("held_out/support_480.jsonl")5agent = SupportTriage(model="claude-opus-5", tools=["crm", "kb"])67harness = Harness(8    dataset=ds,9    scorers=[10        Scorer.exact_match("route"),11        Scorer.rubric("tone", judge="claude-sonnet-5"),12        Scorer.guardrail("pii_egress"),13    ],14    concurrency=16,15)1617# nothing ships that loses to the incumbent on held-out cases18report = harness.run(agent, seed=218)19report.assert_no_regression(20    baseline="prod", metric="route.exact_match"21)

router head · 7 → 10 → 10 → 4

billing41%

technical33%

account19%

escalate7%

recall@50.9112k chunks

cost / case$0.004p50

refusals2 flaggedpii/egress

verdictship+1.8 over prod

Representative evaluation run · not a live system

Systems that do the work, not systems that suggest it.

We ship production AI systems — agents, RAG pipelines, evaluation harnesses, the orchestration around them. No demos that die in QA. Every system we deploy is measurable, observable, and improvable.

01LLM agent orchestration

02Retrieval architectures (RAG, graph)

03Eval & guardrail frameworks

04Fine-tuning & distillation

05Workflow automation (n8n, Temporal)

Next practice

Automation & AI

Systems that do the work, not systems that suggest it.

Custom Application