Vibe Coding a Digital Tim: Part 4 – Rethinking Digital Tim

Digital Tim has been live for a while now, and like most things you build and actually use, the cracks start to show pretty quickly. On the surface, things worked. People could ask questions, Tim would answer them, and everything technically functioned. But once I started looking a bit closer – especially at how the system handled time, memory, and growth – it was clear the foundations needed a rethink.

So this is a walk-through of how I got to Digital Tim 2.0, what was broken, and how the whole thing is now significantly smarter (and easier to live with).

The Original Setup (And Why It Wasn’t Enough)

Every interaction with Digital Tim was being saved to a database table on my NAS. That meant I could open MySQL Workbench and inspect what people were asking.

Technically fine. Practically… not great!

It required hopping between tools, manually poking around tables, and doing mental gymnastics to figure out usage patterns. No overview. No trends. No visibility. Worse still, time-based questions completely fell apart.

Ask something like:

“How long have you been a UX designer?”

And you’d get:

“I started UX in 2018 after switching from development.”

True… but not the “right” answer.

Or:

“When was your last DJ set?”

And Tim turns into a deer with no eyes! (No idea)

The Real Problem: Vector Stores Aren’t Time-Aware

This wasn’t an LLM issue. It wasn’t even a prompt issue. It was Qdrant. Vector stores are brilliant at semantic search.

“What”, “why”, “how”, “tell me about” – all perfect.

But anything involving:

when
how long
most recent
first time

…it hasn’t got a clue!

A vector store doesn’t understand dates. It just understands similarity. So asking “When was your last DJ set?” is meaningless unless there’s a structured place where time actually exists.

The Other Pain: Updating Tim Was a Chore

Adding knowledge to Digital Tim involved:

Creating JSON files by hand
Writing tags manually
Generating summaries myself
Running Node scripts to upsert everything

It worked, but it was clunky, a little error-prone and definitely not something I wanted to keep doing long-term! Blog posts were especially annoying. Every new post meant a whole separate ingestion step, figuring out summaries, tags – the whole thing became cumbersome and tedious!

Step Back and Rethink things

One of the upside of doing a project like this are the learnings you have along the way. Give it enough time with anything that’s kind of new and you’ll have those “Ahhh…I should have done it like this in the first place!” epiphanies! With that being the case…it was time to come up with a new plan.

The new plan had three clear goals:

A proper admin panel
Automatic ingestion of content
Real temporal awareness

The Admin Panel (Finally)

Everything now lives behind an internal admin interface. From one place I can:

See how many conversations happened in the last 7 days or 7 weeks
Inspect what Digital Tim actually knows
Add, edit, or remove memory
Test responses
Take Tim offline
Switch LLM models
Review answer success rates
Review answer response times

No scripts. No guesswork. I’ve now got a ton more visibility and way more control. This bit alone was worth the rebuild!

Digital Tim - Settings

Digital Tim - Dashboard

Digital Tim - Q&A Editor

Digital Tim - Chat tester

Automated Knowledge (No More Manual Ingestion)

I already had the data – I just wasn’t using it properly. My:

Blog
YouTube channel
SoundCloud

…all expose RSS feeds.

So now:

n8n checks each feed hourly
New content is detected automatically
Changes are diffed intelligently
Summaries and tags are generated by an LLM
Everything is written to MySQL and Qdrant

If I publish a blog post, I don’t need to do anything else! Within an hour, Digital Tim knows about it. Boom!! How it should be…

Temporal Awareness (The Big One)

Here’s the key shift.

Qdrant is still used for semantic meaning – the what.
MySQL now handles facts, structure, and time – the when.

All Q&A items exist in both places:

Qdrant for recall
MySQL for structure

SoundCloud, YouTube, blog posts – same deal.

The updated prompt now:

Knows today’s date
Knows which data lives where
Can query MySQL for timestamps and durations
Can do date maths when needed

So now:

“How long have you been a UX designer?”

I properly shifted into UX design in 2018, so I’ve been doing it for about seven years now.

“When was your last DJ set?”

My last DJ set was the ‘Cars Hackathon ’24’ mix, which dropped on 26th July 2024. It was a cheeky house set for the closing party at Booking.com.

Tim finally understands time.

What’s Next?

Reviews! I watch a lot of films and TV. That data is going in next:

Favourite shows
Favourite films
Franchise opinions
Ratings and mini-reviews

The structure is already there – it just needs filling.

Digital Tim 2.0

Digital Tim is now:

Smarter
Time-aware
Easier to update
Largely self-maintaining
Actually observable

Less manual effort. More accuracy. Better answers. Boom!

My boy is growing up.

The Original Setup (And Why It Wasn’t Enough)