AI Workflow Automation.
LLM pipelines and agents that replace the work your team shouldn't own. We design the graph, write the evals, run it in production, and stay on call when it drifts.
What shows up in your repo.
Agent graph
Typed nodes, tool schemas, retry and fallback policy. Versioned like code.
agents/*.ts · dag.yamlEval harness
Golden sets, LLM-as-judge, regression runs on every PR.
evals/*.jsonl · ci.ymlObservability
Traces, token costs, hallucination rates, drift alarms. Wired to your stack.
otel · datadog · honeycombRunbook
What to do when a tool 500s, when latency spikes, when eval red-lines.
docs/runbook.mdEvery arrow is typed. Every node has an eval.
RAG that survives contact with production.
Most RAG demos fall over on real corpora — stale chunks, wrong retrieval, no way to tell when the answer is ungrounded. We build the parts that make it dependable.
Chunking & embeddings
Document-aware chunking, the right embedding model for your domain, and an index you can rebuild on a schedule — not a one-off notebook.
Hybrid retrieval
Vector + keyword + metadata filters, re-ranking, and citations on every answer so a human can verify the source.
Groundedness evals
LLM-as-judge graders for faithfulness and relevance, run nightly and on every PR, so a regression is caught before your users find it.
Freshness & cost
Incremental re-indexing, cache strategy, and token budgets — so the system stays current without the bill running away.
If we can't measure it, we don't ship it.
Every system lands with a golden dataset, LLM-as-judge graders, PII and jailbreak guards, and nightly regression runs in CI. You get a pass/fail number on every pull request.
From a brief to a shipped, evaluated agent.
$ll plan "cut ops triage time by 60%" ▸ trace one week of the target workflow · 2d ▸ design agent graph + tool contracts · 3d ▸ author eval set + graders · 2d$ll ship agent --env=prod --guarded ✓ first guarded production use · week 5$ll evals --since=30d faithfulness 99.4% ops_hrs_saved 1,840 / moA pricing engine, rebuilt in six weeks.
One representative case. Read the full editorial walkthrough, including the deploy log and latency chart.
Rebuilding a pricing engine in six weeks.
Series C fintech · monolith → 4 services + gateway · 42 TB ledger migrated · zero downtime.
READ THE CASE →Six things finance + eng teams want answered.
Discovery (we trace a real workflow before proposing anything), an agent-graph architecture, a typed tool layer over your existing systems, an eval harness, observability, and an operate phase where we stay on call. Not a slide deck — a system in production.
Yes. RAG is one of the most common shapes we ship: document-aware chunking, hybrid retrieval with re-ranking, citations, and groundedness evals so you can trust the answers. We treat it as production software, not a demo.
Claude is our default — we are Claude-native — but the graph is written so the model is a swappable dependency. We also run OpenAI, Google, or open-weights on your infra when the job calls for it.
Groundedness and policy graders run on every change, PII and jailbreak guards sit in the request path, and a runbook defines what happens when a tool 500s or an eval red-lines. Nothing ships without a measured pass rate.
Yes — the router speaks the APIs your team already runs (Salesforce, Zendesk, Stripe, Snowflake, internal services) through typed tool schemas. The agent calls your systems; it does not replace them.
We operate what we ship: SLAs, on-call, nightly evals, and prompt/grader tuning as the world changes. The retainer is tied to eval runs and on-call, reviewed quarterly — no surprise hours.