Built with Anthropic Claude
KUMO builds production software with Anthropic Claude for revenue stage teams. We handle model selection across the Claude family, evals, integrations, cost controls, and handover. CampaignHQ uses Claude in production today.
The Claude ecosystem
The model family we build on, and the product we already operate on Claude in production.
KUMO builds across the current Claude model family: Fable 5 (Mythos-class · long horizon autonomous work), Opus 4.8 (top reasoning), Sonnet 5 (default workhorse), Haiku 4.5 (cost tier). Model tier selection and routing are decided per workflow in the architecture review.
Our practice aligns with Anthropic's Responsible Scaling approach: evals gate every prompt change, and human review runs on every risk tier decision. KUMO participates in the Claude Partner Network process.
Proof from KUMO's own product
CampaignHQ uses Anthropic Claude for selected AI features in production. KUMO operates the workflow, monitors cost and output quality, and controls model changes through its own evaluation and rollout process.
That experience informs how we design evals, cost controls, and handover for client Claude builds.
When custom Claude is right
Four situations where Claude beats the alternative and a custom build is the right call.
Claude's workhorse and top reasoning tiers support long context workflows, subject to retrieval quality and eval results. When your workflow ingests full contracts, medical records, or multi document research, Claude is a strong candidate for maintained coherence across the input.
Claude is a strong candidate for multi step tool use, browser workflows, and agents that need to call your CRM, product APIs, and internal systems. For very long horizon autonomous work, Anthropic's Mythos-class model is designed to run for hours where availability supports it. We validate agentic performance against your specific workflow in the architecture review.
Claude offers refusal and steering behavior that many teams choose for outputs where errors carry material financial, legal, clinical, or customer risk. That behavior is one input to a broader control design that includes policy checks, validation, logging, and human review.
We route tasks by complexity across the Claude family and use prompt caching where context repeats. The cost tier handles simple paths, the workhorse tier covers most production work, and the top reasoning tier is reserved for cases that justify its cost. We validate the mix against your eval set, traffic, latency target, and budget.
What KUMO ships
Agents that call your CRM and product APIs, maintain context across steps, and route by task specific validation to human review. Model selection across the Claude family per step. For very long horizon autonomous tasks, we evaluate the Mythos-class option where availability supports it. This pattern runs across CampaignHQ AI features in production: the workhorse model handles the main workflow, deterministic checks validate the output, and low confidence cases move to review.
Contract analysis, medical record extraction, financial filing review, multi document research. Structured output with schema validation. Human review checkpoints at every risk tier.
RAG on your company data with proper governance. Source citation on answers that need to be traceable. Task specific validators below which the agent defers to a human. Access controls that respect your existing permissions model.
Model routing across the Claude family based on task complexity. Prompt caching for repetitive context. Batch API for asynchronous workloads. Cost attribution per user, per feature, per model.
Production operating model
Every KUMO Claude engagement starts with a common control baseline, then adds the privacy, review, observability, and deployment requirements your workflow needs.
Dev, staging, and production with independent API keys, usage caps, audit logs for decisions that touch customer data.
Task and workflow level evals with regression tests on accuracy, format compliance, and refusal behavior. Model or prompt changes go through eval delta review before ship.
Prompt caching, task based model selection, batch processing where appropriate, cost attribution per user and per feature. Budget alerts and cost dashboards configured before launch.
Task specific validators, eval thresholds, deterministic checks, and human review rules. Outputs that pass checks continue; exceptions move to review.
Data handling scoped per workflow, contract, region, and risk tier. PII controls and access controls designed to match your policy, not assumed.
Every prompt versioned, every eval documented, failure modes and escalation paths written down. The handover is designed so your team can operate the system without a mandatory retainer.
Bring one workflow. Its inputs, its accuracy requirement, its cost ceiling, its latency budget, and its risk tier. Get a 30 minute architecture review. KUMO tells you which tier of the Claude family fits, whether alternatives fit better, and what the production build looks like.
Delivery approach
Map workflow, inputs, accuracy requirement, cost ceiling, latency budget, and risk tier. Model tier selection across the Claude family. Application layer runs on your cloud of choice; Claude routes through the Anthropic supported surface that fits your procurement, region, and compliance needs.
Data quality review. Redaction and PII plan. Eval dataset built from real inputs. Success metrics defined per workflow. Human review checkpoint design.
Working prototype against real data. Eval harness live. Accuracy, refusal behavior, and cost per query measured. Formal go or no go review against agreed success criteria.
Full production integration. Model routing across the Claude family. Prompt caching in place. Validation and fallback wired. Human review checkpoints deployed. Observability dashboards live.
Real cost per user and per feature measured. Prompt caching effectiveness verified. Model routing thresholds tuned against evals. Budget alerts wired to your on call rota.
Runbooks handed over. Your team trained. Optional retainer for prompt tuning, model refresh when Anthropic ships new versions, and new workflow additions.
Typical engagement
Three shapes of engagement based on scope and stage. Anthropic API costs are billed by Anthropic directly. Final quote after the architecture review.
Typical investment
$10K to $25K
4 to 10 weeks
One narrow Claude workflow shipped end to end. Single agent or single document processor, eval harness, prompt caching, validation and fallback, human review checkpoint, runbooks, handover.
Typical investment
$25K to $50K
10 to 20 weeks
Multi agent Claude system with model routing across tiers, deep tool integration, cross workflow observability, migration from a hosted chatbot platform, or compliance sensitive setup with Bedrock or Vertex.
Typical investment
$50K to $100K / year
Annual, ongoing partnership
Ongoing multi project engagement for teams running Claude at high volume. Continuous prompt tuning, model refresh when Anthropic ships new versions across the family, new workflow additions, incident response, cost engineering, cross region expansion. Renews annually.
What changes the range: model mix across the Claude family, workflow complexity, integration depth, data volume, compliance and residency requirements, and how much of the eval harness is greenfield. Anthropic API costs are billed by Anthropic directly. Final quote after the architecture review.
Why KUMO fits
A senior, hands on team for 4 to 20 week production builds, with direct access to the people doing the work and full IP transfer at handover. Best for teams that want serious delivery controls without a multi vendor transformation program.
Prompt caching, task based model selection across the Claude family, batch processing where appropriate, and usage attribution are designed into the system. Before launch, we define cost per successful task, budget alerts, and the reporting your team needs.
Task and workflow level evals with regression tests on accuracy, format compliance, and refusal behavior. Model and prompt changes are measurable and reversible, including when Anthropic ships new versions.
This experience comes from operating KUMO's own product in production. CampaignHQ uses Claude for selected AI features today, and every client build inherits what we learned.
Related KUMO work
FAQ
Three things. We use Claude in production for CampaignHQ AI features, so the patterns we ship for clients are the same patterns we run for ourselves. Cost controls and eval harnesses are designed in from day one, not added later. And you get a senior team from architecture review to handover, with full IP transfer.
We build across the current Claude family: Fable 5 (Mythos-class, for long horizon autonomous work, subject to current availability), Opus 4.8 (top reasoning for the hardest cases), Sonnet 5 (default workhorse for most production workflows), and Haiku 4.5 (cost tier for high volume paths). Model selection and routing are decided per workflow in the architecture review, and we validate on your eval harness before switching production traffic when Anthropic ships new versions.
Depends on the workflow. Claude has reported strengths in agentic tool use, safety oriented steering, and long context reasoning. OpenAI's GPT family has the largest tool ecosystem and strong terminal coding. Google's Gemini family is well suited to multimodal workflows and Google Cloud aligned stacks. We compare candidates against your actual evaluation dataset in the architecture review rather than relying on marketing benchmarks. If another model is a better fit for your accuracy target, latency budget, or ecosystem, we say so.
Depends heavily on model mix across the Claude family, volume, prompt caching effectiveness, and batch eligibility. Prompt caching can reduce cost meaningfully on cacheable workloads. Batch processing offers additional savings for asynchronous workloads. We model your specific workload during architecture review with cache hit assumptions, model mix, and volume, and ship cost dashboards so you always know per user and per feature spend.
Depends on procurement, region, cloud commitments, feature availability, and contract terms. Anthropic API gives the shortest path to the latest features. Bedrock and Vertex fit teams with existing cloud commitments, procurement requirements, or specific data residency needs. Model and feature availability differs by provider and region, including which Claude tier is available where. We evaluate the choice against your specific constraints in the architecture review.
We define data handling per workflow, contract, region, and risk tier. Anthropic states that commercial API inputs and outputs are not used for model training by default, subject to opt in and contract terms. Where procurement or cloud requirements call for it, we evaluate Anthropic's direct API, Amazon Bedrock, and Google Cloud Vertex AI based on current model availability, regions, features, and data terms. Redaction, access controls, logging, retention, and human review are scoped explicitly rather than assumed.
Most workflows improve first through better context, retrieval, tools, structured outputs, and an eval set built from real examples. KUMO tests those options before considering any provider specific customization feature. Current fine tuning availability varies by model and platform, so it is verified during architecture review rather than promised by default.
Full IP transfer at handover for the custom code, prompts, eval datasets, and runbooks KUMO builds. Anthropic's Claude models remain Anthropic's. CampaignHQ remains KUMO's product. Everything KUMO writes for your project is yours.
Depends on workflow differentiation, integration depth, risk tier, ownership requirements, and total operating cost. If an off the shelf product covers 80 percent of the workflow at acceptable quality and cost, we say so. If the workflow needs deeper integration, custom logic, or specific control patterns that off the shelf products cannot deliver, a custom build is often the right call. We take you through this comparison in the architecture review.
30 minutes, no pitch deck. You describe one workflow. Its inputs, its accuracy requirement, its cost ceiling, its latency budget, its risk tier. KUMO tells you which tier of the Claude family fits, whether alternatives fit better, and what the production build looks like.
