Workflow definition
A defined objective, boundaries, tool contract, state, escalation path, and human handoff before implementation.
AI INTEGRATION FOR EXISTING SOFTWARE TEAMS
AI Product Integration
Your product already has users and a workflow worth improving. I design and integrate one AI feature, or take an existing AI prototype through the reliability work needed for release.
Scope your featureBUILT FORSoftware teams adding speech, document search, or an assistantProduct teams with a defined user workflow and representative inputsCompanies that need private knowledge retrieval and controlled tool useExisting AI prototypes that need testing, reliability, and monitoring
WHAT IS INCLUDED
From integration
through release.
Tools and integrations
APIs, databases, documents, internal services, browser actions, MCP servers, and approval checkpoints.
Knowledge and memory
RAG, structured context, retrieval quality, permissions, and memory designed for the actual domain.
Test cases and evaluations
Representative cases, pass criteria, failure analysis, regression checks, and staged permissions.
Security and control
Least-privilege access, audit trails, cost limits, safe retries, and human approval for consequential actions.
Monitoring and handoff
Deployment, traces, model routing, latency and cost tracking, iteration, and written system documentation.
HOW THE ENGAGEMENT WORKS
One feature. An agreed definition of done.
Start with a product walkthrough and sample inputs. I propose a bounded scope, timeline, and fee before implementation. The sprint includes integration, representative evaluations, release support, and a written handoff; ongoing maintenance is scoped separately.
Define the job and the boundary
Use real examples to specify the trigger, inputs, systems involved, success criteria, failure cases, and where a person must remain in control.
Connect context and tools
Add retrieval, structured context, and narrow tool contracts with least-privilege access, explicit state, cost limits, and safe retries.
Evaluate before expanding
Run representative cases, inspect failures, measure quality and latency, and widen permissions only when the workflow earns that trust.
STARTING POINTUseful first input: ten representative examples, the current manual workflow, the systems involved, and the decision a human makes today.
TECHNOLOGY
Selected for the app.
OpenAIAnthropicMCPRAGTypeScriptPythonCloudflare
RELATED PRODUCT WORKMurmur MCP toolsAudioPage document chatLocal speech and retrieval products
RELATED WORK AND NOTES
Apps and engineering
decisions.
INDEPENDENT PRODUCT · macOS
Murmur
A private voice studio for Mac.
INDEPENDENT PRODUCT · iOSAudioPage
Read less. Listen deeply.
INDEPENDENT PRODUCT · macOSEchoText
Speak. It types. Anywhere.
ENGINEERING NOTE · 9 minHow to scope an AI workflow agent with permissions and evals
Define the task, permitted actions, test cases, and human review before choosing how much the system can do on its own.
COMMON QUESTIONS
Scope and delivery.
Do we need an existing product?+
This offer is for teams with an existing product or working prototype, a defined user workflow, and access to the code and data needed for integration. Bring representative examples so we can agree on what a successful feature must do.
How are price and timeline agreed?+
After reviewing your product, sample inputs, and integration constraints, I propose a scoped fee, timeline, acceptance checks, and handoff. Model usage and third-party services are identified separately. Ongoing support is a separate scope.
Do you train a custom model for every agent?+
Usually not. Most production value comes from the right workflow, context, tools, permissions, and evaluation. Fine-tuning or adapters are used only when testing shows they are justified.
Can you add AI to an existing app?+
Yes. The work can be a focused product feature such as transcription, document analysis, search, generation, or an assistant. I can use local models, cloud APIs, or a hybrid approach.
What is a good first agent project?+
A frequent, bounded workflow with accessible inputs, a clear human owner, and a result you can check. Examples include document intake, support classification, research, or report preparation.
Will the agent run fully autonomously?+
I begin with observation or drafted actions, keep a human approval step, and expand permissions only after the workflow performs reliably on representative cases.
DISCUSS THE PROJECT
Bring your product.
Define the next feature.
I’ll review the starting point and tell you whether the work fits the studio.
Scope your feature