← Back to home
Service

Custom AI agents for workflows that need their own system, not another layer on top of the old one

This is for workflows where adding AI to the existing stack is not enough. The work is spread across portals, PDFs, inboxes, spreadsheets, browser tabs, and half-manual handoffs. What is missing is not one more integration. It's a system that actually owns the job.

I build that system as a custom backend workflow or AI agent. Sometimes it needs a thin operator screen. A lot of the time it doesn't. If the job is a report that shows up on schedule, an underwriting pass assembled from messy inputs, or a decision package ready for review, that's the product.

Based in

Los Angeles, CA

Best for

B2B founders, operators, and investor-backed teams that need the technical decisions thought through, not just executed.

Outcome

A system you can actually run in production, not a one-off demo that collapses under real usage.

When this is the right fit
  • The workflow needs a dedicated execution layer, not just AI dropped into the tools you already have.
  • Someone on the team has quietly become the operating system, moving context between portals, files, inboxes, and tools that were never meant to work together.
  • The output needs judgment, synthesis, or multi-step execution, so basic automation rules are not enough.
  • You care more about one system owning the job end to end than about shipping another interface for people to click through.
What I build
  • Custom AI agents and scheduled workflows that gather inputs, reason through the job, and deliver the output without someone babysitting a queue.
  • Backend orchestration across APIs, documents, websites, browser automation, email, databases, and whatever ugly source the process depends on.
  • Review steps where a person should approve the draft, the decision, or the send before anything moves.
  • Retries, dedupe, logging, alerts, and cost visibility so the system still works after the demo.
What I care about in production
  • Model output doesn't write or send anything until it passes the checks the workflow actually needs.
  • Secrets stay server-side, and every expensive or sensitive path is owned by the backend, not the browser.
  • The workflow is built for retries and partial failure, because source systems break and webhooks get delivered twice.
  • There is always a fallback path when a portal changes, a login expires, or the model response just is not good enough.
How I usually work

I start by asking whether this needs an integration or its own system. If the workflow needs a real owner, I build the smallest custom agent or backend engine that can take that role in production. The goal is not a clever demo. The goal is that the work keeps getting done when nobody is watching.

Relevant work

LA Market Report Agent

Seven sources. Eight parallel research tasks. One monthly report that just shows up on schedule.

In progress

FunnelScout

Webhooks return fast, background jobs do the heavy work, and AI cost is tracked while the run is still in motion.

View case study →

Dotty

Reminder drafts stop at review. Nothing sends itself because the operator still owns the call.

View case study →
Next step

If this sounds like the shape of the problem, send me what the workflow is doing today, where it breaks, and what has to stay true in production. I care less about the pitch and more about the constraint.