AgentCrafts

Work that replaced a chore people used to do by hand.

Three projects, written up honestly: what was broken, what the agent actually does, and what changed afterwards. Client names are left out; the details are real.

Revenue operations

Inbox-to-CRM agent

Inbound email gets read, the sender enriched, the CRM updated, and a reply drafted for a human to approve or rewrite.

LangGraphNext.jsPostgresSlack
01 / Problem

What was broken

Every enquiry arrived in a shared inbox and sat there. Nothing reached the CRM unless someone remembered, and follow-ups depended on who was at their desk.

02 / Approach

What the agent does

An agent watches the inbox, classifies the message, looks up the contact, enriches the company, then writes the record and drafts a reply. Anything above a value threshold waits for approval in Slack rather than sending.

03 / Result

What changed

The inbox went from a queue people avoided to a review screen. Records land in the CRM the same day and nothing sends without a person pressing the button.

Content operations

Research crew

Four specialised agents brief, draft, check and format long-form research, with sources cited throughout.

CrewAIpgvectorMCPTypeScript
01 / Problem

What was broken

Long reports took days of switching between tabs. Drafts were inconsistent and the checking step was always the first thing skipped when a deadline moved.

02 / Approach

What the agent does

A supervisor agent splits the brief into sections. A researcher gathers sources, a writer drafts, a fact-checker verifies claims against the source text and returns anything it cannot confirm, and a formatter assembles the document.

03 / Result

What changed

A multi-day cycle became a supervised few-hour job. Unverified claims now come back flagged instead of quietly landing in the final draft.

Internal support

Operations copilot

An assistant over thousands of internal documents that answers in Slack and shows where each answer came from.

pgvectorOpenAISlackLangfuse
01 / Problem

What was broken

Answers existed somewhere, but finding them meant searching a wiki, two drives and asking the one person who remembered.

02 / Approach

What the agent does

Documents are ingested nightly with hybrid vector and keyword retrieval, so exact part numbers and phrasing both work. Answers always carry a link to the source passage, and an evaluation set runs on every change to catch retrieval regressions.

03 / Result

What changed

One place to ask instead of six to search. Replies include the source, so people can verify instead of taking it on faith.

A note on this page

I keep client names and internal numbers off the site by default. If you want to see the code, the traces or a live walkthrough of any of these, ask on a call and I'll share what I'm allowed to.

Request a walkthrough

About the work

Is the portfolio real, and how did it go

Why aren't the clients named?

Because most of this work sits inside someone's operations, and a lot of it touches customer data or internal process. Naming clients and quoting numbers would mean asking them to be marketing material, which is not a fair trade for them. The problems and outcomes are described as they happened.

Can I see the code or a live demo?

For client work, usually not, since the code belongs to them. On a call I can walk through the architecture and, where it fits, show something comparable running. Judging whether I can do the work should not depend on seeing a stranger's login screen.

How long did each of these take?

Roughly two to eight weeks, depending on how many systems were involved and how much of the surrounding workflow had to be built as well. The pattern is nearly always the same: something working in the first or second week, then hardening it.

Do you take on work like these?

These three are typical rather than exceptional: document heavy operations, multi step processing, and workflows where someone re types the same information between tools. If your problem looks like one of them, it is worth a conversation.

What happened after launch?

Real traffic found things the demo never showed, as it always does. The first weeks are spent tuning retrieval and prompts, trimming cost and adding the exceptions nobody mentioned at the start. Launch is the beginning of the useful part, not the end of the project.

Let's build

Worth a thirty minute call?

Tell me what your team does by hand every week. I'll tell you honestly whether an agent is the right answer, what it would take, and roughly what it costs to run.