Consultation
Work

Really Focus

Role
Founder, product and engineering
Period
2022 – 2026
Status
Live, no longer in active development
Technologies
Laravel, React, PostgreSQL with pgvector, Azure AI Foundry, MCP, Telegram

As a consultant, I was overwhelmed by email.

Really Focus is the AI chief of staff I built to fix that. It reads the noise, remembers every client promise, catches the time I forgot to bill, and tells me what matters this week.

1 · your inbox, monday 8am472 · really focus sorts itReally Focuspromisebudget to Sarahbillable1.5h with the CTOoverdueinvoice #442follow upAcme proposal3 · your week, sortedSunday briefing1big thing3that matter5to clear£450 invoice drafted1 · your inbox, monday 8am472 · really focus sorts itReally Focuspromisebudget to Sarahbillable1.5h with the CTOoverdueinvoice #442follow upAcme proposal3 · your week, sortedSunday briefing1big thing3that matter5to clear£450 invoice drafted
Really Focus reads the noise, pulls out the promises, billable time and overdue invoices, and hands you a plan for the week.

Admin was eating the actual work

Independent consultants sell their time and their judgement, but a surprising amount of both goes on admin. Monday mornings went on piecing together what happened last week. A quick call that should have been billed never made it onto an invoice. A follow-up promised three weeks ago quietly slipped, and with it a little of the client's trust.

General-purpose AI chat tools didn't help. They forget everything between conversations, they know nothing about your clients, and they wait to be asked. I needed something that worked in the background and came to me.

What changes with Really Focus

Nothing slips through

Really Focus remembers every meeting, email and conversation you share with it. Ask "what did I last promise Sarah?" and it answers from months of history, then nudges you before the deadline passes.

Every week starts with a plan

Every Sunday a briefing arrives, account by account: one big thing, three that matter, and five to clear. Monday morning is for work, not reconstruction.

You get paid for all of your work

It pulls meetings from your calendar, flags the ones that look billable, and drafts invoices in Xero or FreeAgent for you to check and send. It chases overdue ones too.

It works where you already are

You talk to it through Telegram, including voice notes, and it connects to the tools you already use: calendars, Gmail, accounting software, Notion, Linear and more. Claude and Cursor can query it directly too.

How I built it

  • A memory that forgets the right things. Email, calendar and billing data are stored as searchable embeddings, but remembering everything is as unhelpful as remembering nothing. Facts persist, snapshots fade, events fade only when they're ignored, and an AI agent tidies the memory every week.
  • Proactive, but careful. The assistant runs on a schedule in each user's time zone, makes structured decisions, and only acts when it is confident enough to.
  • AI costs kept under control. Fourteen specialised agents each use the cheapest model that does their job well, only the tools relevant to a request are sent to the model, and every call is costed against a monthly allowance.
  • Trusted with sensitive data. All AI runs inside Azure and never trains on user data. Anything that writes to another system needs explicit consent and is audit-logged, and code the AI writes runs in isolated sandboxes with no internet access.
  • Pragmatic about compliance. Gmail access uses each user's own Google credentials, which avoided a security audit costing tens of thousands of pounds without weakening anyone's protection.

The current version was a ground-up rebuild in 2026: around 380 commits, 260 test files and 20 architecture decision records in four months.

Where it is now

Really Focus is still live at reallyfocus.com, but I've stepped back from active development to focus on Spektral. What I learned here about memory, cost control and earning trust with sensitive data carried straight over.

What this means for your project

If you're building an AI assistant or agent that works with real business data, the hard parts are rarely the model. They're memory, cost, integrations and trust. I've built all of those into a working product and can help you get them right. Get in touch to talk it through.