Real work, real outcomes
Each project started with a specific operational problem—inventory chaos, manual dispatch overhead, compliance drift. We mapped where humans were grinding through repetitive work, then built an agent with exactly the right autonomy guardrails. Here's what shipped and what it changed.
Inventory Sync Agent
A mid-size retailer selling across Shopify, Amazon, and Walmart was drowning in inventory drift. Stock levels disagreed between platforms every day. Overselling, stockouts, refunds piling up.
The problem
Shopify said 45 units in stock. Amazon said 62. Walmart hadn't synced in 8 hours. Coordinating across three APIs manually wasn't working—and when stock ran out, customers got refunds instead of fulfillment.
What we built
A SiteCtrl Agent with native connectors to all three marketplaces. It polls inventory every 15 minutes, reconciles stock levels against a shared source of truth in the retailer's warehouse system, and auto-pauses listings the moment a discrepancy appears. No guessing, no delay.
The thinking
We chose polling over webhooks. Marketplace webhooks are unreliable—sometimes they batch, sometimes they don't fire at all—and when inventory is on the line, unreliable is unacceptable. Polling every 15 minutes is predictable and catches drift fast enough to matter.
Outcome
Dispatch Console
Dispatch coordinators were juggling two systems: a legacy TMS (Transportation Management System) with planned routes and live driver check-ins from the field. Reconciling them took hours. Missing a conflict meant a driver showed up at the wrong place.
The problem
Every shift, someone had to manually cross-check TMS assignments against GPS pings and driver messages. "Driver is 20 minutes late to pickup." "Delivery address on file doesn't match what's in the truck." These exceptions were real but invisible until someone looked. The workload was killing productivity.
What we built
A SiteCtrl Console ingesting live TMS data and driver GPS telemetry. It surfaces exceptions in real time—late arrivals, address mismatches, route conflicts—with suggested resolutions pre-filled. One click to approve, and the driver's phone updates. Still human-driven; the agent never auto-dispatches.
The thinking
Dispatch is too high-stakes for full autonomy. A wrong routing call costs customer trust and fuel. We gave the agent a read-and-flag role: it sees everything, surface problems fast, and pre-fills the right answer. The coordinator still makes the call. This trades some automation for confidence.
Outcome
Patient Portal Automation
A regional clinic network was burning hours on no-shows. Staff called patients manually to confirm appointments. Intake forms sat unsigned. Rescheduling was first-come, first-served and chaotic. The no-show rate was 22%—well above industry baseline.
The problem
Patient communication is the blocking task. Reminders, form follow-up, rescheduling suggestions—all require knowing HIPAA rules and talking like the clinic, not a bot. Staff couldn't scale it without adding headcount, and no third-party tool was fine-grained enough to trust with patient data.
What we built
A HIPAA-aware SiteCtrl Agent handling appointment reminders, intake form triage, and rescheduling suggestions. Every patient-facing message is scored against a compliance ruleset before sending. Full audit trail for every interaction. The clinic reviews flagged messages in a dashboard before they reach patients.
The thinking
Patient safety and compliance aren't negotiable, so we built guardrails into the agent itself. It knows what it can and cannot say. It knows not to make medical claims. It knows to flag edge cases for a human. This isn't about restricting the agent—it's about being honest about where autonomy works and where it doesn't.
Outcome
Site Ops Agent
A law firm's website is a compliance document. Bar-required disclosures, malpractice insurance links, attorney bio accuracy—all have to be right or the firm risks sanctions. But the site was never audited consistently. Broken links and outdated disclaimers piled up.
The problem
No one had time for nightly audits. A state bar exam is not the place to discover your disclosure links are dead. The firm needed accountability but had no operational process to create it.
What we built
An always-on SiteCtrl Agent that crawls the site every night, checking for required disclosures, validating links, and comparing attorney bios against the bar's official roster. Any discrepancy files a ticket automatically. The agent has zero write access to the live site—it only reads and reports.
The thinking
Read-only agents are low-risk. They can't break anything. That's valuable for compliance work where the worst move is changing something you shouldn't. The agent catches problems early; the firm decides what to do. This is scalable auditing, not automation.
Outcome
The pattern
Every one of these starts the same way: we map where a human is doing repetitive, rules-based work on or around a website. Then we ask a harder question: how much autonomy should the agent have?
Read-only? (Almeida & Cole: auditing, no write access.) Flag-and-approve? (Fenwick: surface exceptions, human decides.) Full autonomy within guardrails? (Northwind: auto-pause listings, but only when stock is truly wrong.) The shape of the agent follows the answer, not the other way around.
That thinking is the difference between an agent that works and one that fails. See our process for how we get it right before we ship.