Building Duel-Engine AI Assistance that get work done

• AI & Automation
  AI & Automation

  From automated workflows to real-time assistants, I’m learning to build tools so you can stay focused on what really matters.

  Learn More View Articles

Please Note: This article was not written or thought by AI. AI is essential for many things, but authentic knowledge and opinions is essential for real insights, and the opinions of those working within the industry.

I used to dream about an assistant that could understand our day better than we do. This assistant would decline a lunch meeting because it knows we’ve skipped breakfast, or flag an urgent message buried in chat like Slack – or even reply to a client in a tone that sounds like us.

Well, that dream no longer requires a team of data scientists or a warehouse of GPUs. It just needs to brains working together: a lightweight local language model for speed, and OpenAI’s API for depth of knowledge and enhanced accuracy.

This is an article about how I sticked the two together, and why I think its the missing piece of personal in AI.

Why two brains, instead of one?
Duel Language Models.

Local Language Model

The local model sits on your device and the latency drops to near-zero – or in other words – its instant, like a sprinter.. it skims long documents and shrinks them in to summaries. Its also able to keep up in realtime with calendars, reminders, browser tabs, chats and other useful information operating on your machine.

Large Language Model

Open AI’s model is the master chess player, its a strategist. I feed it with distilled snippets and it can return richer reasoning and better phrasing. Its suggestions are better – and it contains smarter scheduling logic. Letting the Local LM clear the unnecessary data the OpenAI large language models can focus on the insight instead of the initial grunt work – tasks you’d be happy for a slight delay to produce an accurate response.

The end result is that responses can feel instant but can still carry the power that only a large language model can deliver.

Reading the room (your desktop) with AI

Using AI models, we can give permission to gather data from many different aspects of a workflow – or applications on a machine.

Calendar events
Time-tracking logs
Teams/Slack channels
Email threads
Active browser tabs

Using this data, AI can answer those essential automation questions, “When can we meet?” and schedule in a meeting that respects working hours and other scheduling conflicts – it’ll even leave you with breathing space between the scheduled calls. We can also put guardrails in place, like maximum meetings per day or no bookings after 5pm as a general rule. We can build these as code checks that can manipulate the response going to or from AI. This ensures no conflicts, and can re-request a time from AI – before the calendar event is added.

One voice, many faces. Multi AI Personas.

Not every message should sound the same, and we can feed AI training profiles from past conversations. The assistant learns how you talk to:

Long-term clients (warm and concise)
A new lead (Small detail, enthusiastic)
Your own team (To the point, with emojis)

The local model tags incoming requests with the right persona and the OpenAI model handles the final wording. The replies feel handcrafted and tailored, whilst maintaining previous conversational details.

And we can go further than that, each individual team member of your own team will have bespoke replies – AI will be aware if they are in to football, or what projects are being collaborated on.

The AI Demo

I wired calendars, mail, reminders and chat in to a single React application. the local language model watches for context changes – and when you get pinged, the current task timer can stop, the chat window is focused and the incoming message is analyzed. The packet goes to Open AI which drafts the reply and proposes the next steps.

Try it and you’ll see automation that feels like magic. A team mate sends the message “Free for a quick call?” and a few seconts later the assistant answers, offers two times that won’t clash with other events in your afternoon – and logs a reminder to notes, You’d swear a real person responded and scheduled in th the meeting.

2022 Predictions, 2025 Reality.

In 2022, I write that AI would move from generic chatbots to context-rich and deeply personal tools. Three years later the duel-engine approach proves it. It feels closer to what Apple was hinting at with “Apple Intelligence”, but with the freedom to run on any device and speak in your “voice”.

Reclaim your day.

I build platforms that slot this tech in to real workflows. No mystery, just clean code and tangible time saved. If you are curious about turning your scattered data in to a single quietly brilliant assistant, lets talk.

The best automation isn’t something that will replace you, it shows up when it matters, helps with efficiencies and is mostly invisible. It’ll vanish when its not required, and should be seen to be invisible.