Snapfast
I designed an agent that reads real posts in a creator’s circle and recommends the next one. The creator still publishes.
A daily brief tied to real posts.
Product development
An AI agent on one real job, proven with users before you fund the platform.
For teams with a repetitive job and the people to own it, who want a working proof before a platform build.
Email me
Tell me what’s broken and where you’re stuck.
What you get
A working proof, early
Stakeholders and users react to a real prototype before a full build is on the table.
An agent with a job
Goals, tools, guardrails, and handoffs are defined around one real task. The teams that make AI work are obsessed with the workflow, not the model: some steps get an LLM, some get a plain rule, and some stay with a person.
People stay in control
Approvals sit on the steps that carry risk, not on every action. Ask a person to confirm five times a minute and they confirm everything. States and handoffs are designed so the team knows what the system did and trusts it enough to use it.
Pass/fail before production
Before anyone writes a prompt, we collect 20 to 100 real cases of the job done right and score the agent against them at every step. Title Resource Group automates a step only after 99 of 100 examiner approvals. No vibe checks.
A stack your team can run
APIs and current AI tools, chosen to fit your environment and stay maintainable after handoff.
Specs engineering can extend
When the prototype holds up, you get requirements, UX specs, and front-end your team can ship and maintain.
Examples of this work
Title Resource Group
I put AI on document sorting, exam codes, and boundary sketches. A step is fully automated only after 99 of 100 examiner approvals.
Already 30% of the annual cost target.
BC Tool
I prototyped the workspace people could use: upload a wave, check it, publish the results.
Four years of brand tracking, side by side.
How it works
- 01
Pick one job
We choose a single task the agent or workflow must complete. A demo with no job is out of scope.
- 02
Design the loop
Goals, tools, guardrails, and the handoff back to a person, plus a reference set of real cases done right so there is something to score against. The human stays in control until the data says otherwise.
- 03
Build a proof
A clickable or working prototype your stakeholders and users can react to. This is the test, not the platform.
- 04
Handoff or ship
If it holds up, you get specs and front-end your team can extend. If it does not, you stop before the expensive build.
The approach
One job, a reference set, and a bar to clear.
An agent is non-deterministic software, so it needs a different lifecycle: scope one capability, build the reference data, evaluate every step, and widen autonomy only when the score holds. This is the loop from the AI product guides in Lenny Rachitsky's archive, and it is how Title Resource Group went from prototype to production in three weeks.
- Scope
- One job with a clear input and output. Version one is the narrowest useful slice: route the document, draft the summary, flag the exception. The platform comes after the slice earns it.
- Reference set
- 20 to 100 real cases of the job done well, pulled from logs or built with the people who do it today. It breaks the cold start and tells us what context the agent needs.
- Evals
- Pass/fail scores per step against the reference set, not a demo that worked once. Where the score drops is where the design changes: a guardrail, a narrower input, or a handoff to a person.
- Earned autonomy
- The agent proposes, a person approves, and a step runs on its own only after the approvals say it can. At Title Resource Group that bar is 99 of 100.
Fit
This is for
A real workflow, one agent
You have a repetitive job, an existing team, and you want a prototype people can trust before you scale.
This is not for
ChatGPT on the homepage
Not a fit if you want a generic chatbot with no workflow, or a full platform build with no prototype in between.
FAQ
Both, when the prototype earns it. Title Resource Group went from a front-end prototype to a production tool in three weeks. Most work starts as a proof so you can stop early if it fails.
We agree on the bar before we build: a reference set of real cases and a pass rate per step. The prototype goes to the team when it clears the bar on the cases that matter, and a step runs unattended only when the approval record says it can. That is what Title Resource Group's 99 of 100 means in practice.
Whatever fits your environment: current LLM APIs, the tools your engineers already run, and front-end they can maintain after I leave. I will not lock you into a stack you cannot run.
Even better. I design the agent, the UX, and the handoff. Your team extends and owns it. I am not here to replace an engineering org.
Email me the job you want automated and who does it today. I will tell you if it is a two-week proof or a longer build.