kai ellenson.

The problem

A high-value sale never happens in one message. It takes context, timing, the right offer, and knowing when to follow up.

What we built

An engine that reads the whole conversation, picks from real offers at real prices, sends the message, and schedules the next move. Routine conversations run on their own. Anything unusual goes to a person.

More than a chatbot

The AI makes the call. The system carries it out: which offer, when to send it, when to follow up, and a record of what actually happened. The conversation, the pricing, and the delivery all work as one.

Under the hood

Deciding and sending are separate steps. Before anything goes out, the system re-checks its permissions, the latest state of the conversation, the offer, and the price. Each conversation is handled in order, with delivery receipts and a full history. If a send can't be confirmed, it's recorded as unconfirmed, so nothing gets blindly sent twice.

How we test it

A replay environment runs the AI against past conversations without sending anything live. Every new version is compared against the last one, and release checks flag what got worse before it ships. Every decision keeps the version and the context that produced it.

The problem

Running a company means knowing who owns what, what's actually happening, and how it all ties back to the numbers. That only works when every department is looking at the same picture.

What we built

A custom platform that brings it all together: accounts, teams and departments, training roadmaps, payments and invoicing, shared files, and analytics.

Under the hood

People see what their role and department should see, and nothing else. The same records feed every workflow and every report, so everyone is looking at the same numbers. Heavy work runs in the background instead of making anyone wait on a page.

The problem

Generating an image or a video takes time, and a conversation can't freeze while it happens. Meanwhile, who can see what, and what it cost, has to stay right.

What we built

A platform with live conversations, AI replies, image and video generation, media libraries, and a credit system that records usage.

Under the hood

Generation runs in the background while the screen shows progress. When it finishes, the result is saved, the usage is recorded, and the conversation updates live. Private media is served through signed links, people can only read their own conversations, and every balance change leaves a transaction record.

The problem

A routine update and someone wiping out history can look almost the same at a glance. The team needs to know which changes deserve a second look.

What we built

A GitHub app that watches activity across our repositories, compares the history before and after, checks certain code and dependency changes, and alerts the team by severity.

Under the hood

It connects related events over time, so deleting a branch and recreating it doesn't pass as two harmless actions. Repeat alerts are muted, but new evidence triggers a fresh one. An alert only counts as delivered once it's actually accepted. It's focused monitoring, built to catch specific things well. It doesn't pretend to be a full security audit.

The problem

Finding a page is only the start. Results can be irrelevant, duplicated, or too thin to trust. Someone still has to work out what's real and what to do about it.

What we built

A pipeline that runs in the background. It finds candidate pages, classifies them, scores the evidence, and turns what survives into review and follow-up tasks.

Under the hood

Cheap checks run first, so the expensive work only happens on pages that earn it. Heavy network jobs are capped so they can't pile up. Every decision keeps a snapshot of the evidence it was based on, and a finding is never treated as proven just because a search returned it.