Daniel Parke

Senior Data Scientist working on Energy Optimisation & Grid Trading

PhD student by day, energy wizard by night 🧙‍♂️

[Daniel Parke @DanielParke_]

I was working on an energy model to explore Solar PV variation over a year☀️

I made a quick animation to visualise the data, it turned out pretty well, so thought I would share it 📊

Model was for three arrays totalling 27.3kWp, mostly East/West facing

#SolarPower #energy

[Daniel Parke @DanielParke_]

@noctus91 @UnslothAI @googlegemma https://t.co/AXwW7l70jh

[Daniel Parke @DanielParke_]

Couldn't wait to finish Part 2 of the budget build and decided to give the newly released Gemma 4 12B model a try.

Model was run locally using llama.cpp on a 3090Ti, with a power limit of 350W/450W, got an average speed of 72 tk/s.

Asked a relatively complex question around trading in the UK energy markets, which it handled pretty well.

Really promising results so far, both in terms of quality and speed!

Model page:
https://t.co/6EEhK1Zin0

[Daniel Parke @DanielParke_]

@atomic_chat_hq Nice work, that's a really awesome way to visualise the differences in output quality!

[Daniel Parke @DanielParke_]

First video of many released focused on the Local AI Budget Build project 🤖

Part 1 is the build guide for our initial ~£1k platform, where I walk you through the entire build process, and lay out the long term goals of the project ⚡💵🧮

I am not an expert by any means in this area, so if I can do it, you most certainly can too 🤓

I have also put together a BOM & Build guide document for the initial platform, which can be found in the project Github repository (see video description for links) 🔗

Part 2 will be released in the next week, where we will take you from a fresh Linux Mint install, to having Unsloth/Llama.cpp deployed and available on your local network 🛜

I used RaveOS back in the day to get into crypto mining when the sight of the CLI still terrified me, and this long term project is my way of paying homage to their efforts, and trying to emulate the same obliteration of entry requirements to local AI, that they achieved for sophisticated PoW mining ⛏️

As always, any feedback at all is very much appreciated, and please do not hesitate to reach out with any questions! 🙂

https://t.co/q0tLN1JX6M

[Daniel Parke @DanielParke_]

@AIVibeCoding Thank you! The first of many to come I hope! :)

[Daniel Parke @DanielParke_]

New release for the Hermes Agent Control Hub dashboard 🤖

I hadn't planned on making any more changes, but I set up a few recurring missions and was pretty pleased with the changes, so here they are!

- UI/UX redesign & overall performance improvement
- Helper functions created throughout application
- Hardened sync actions from Hermes Agent -> Control Hub DB
- Overhaul sessions page to increase usefulness
- Many bugfixes and improvements

For the next while I will be running a test using recurring missions to allow the Agent(s) to review and refactor the existing codebase, with minimal intervention.

So you should see an increase of updates to our dev branch.

If you would like to try out these changes, you can sync with the dev branch, and then back to main at anytime from the dashboard (see screenshot below).

As always feedback is very much appreciated, and don't hesitate to reach out!🙂

https://t.co/kod35fI89o

[Daniel Parke @DanielParke_]

It's alive!!! 🤖

Early testing of the budget build has gone well, with promising results 🎉 As such, we have expanded to include some more GPUs for testing.

Total build cost:
RTX 3090Ti (24GB) - £1149
RX 7900 XTX (24GB) - £1009
RTX 2070 super (8GB) - £484
NucBox G5 - £145.95

Early tests with tuning showing power reductions of around 33%, while getting token speeds anywhere from 50-200 tk/s depending on the model.

This is using a range of models, mainly the Dense/MOE Qwen-3.6 & Gemma-4 models.

Video build guide + install script to get started with Llama.cpp/Unsloth coming soon!

[Daniel Parke @DanielParke_]

New project on the way to see what is possible using local AI, with a setup that is actually affordable and possible to build right now.

Project goals & restrictions:
- Budget Scale: Can a viable 24GB VRAM AI rig be built for ~£1k, and a dual-card 48GB rig for < £2k?
- Power Efficiency: How much can GPU tuning reduce power demand while keeping performance acceptable?
- AMD vs. NVIDIA: Can identical power and kernel-level optimisations be achieved across both architectures?
- Grid-Aware AI: Can the rig dynamically alter its tuning profiles in response to energy schedules or real-time load caps?

For this project, "viable" means maintaining at least 25 tokens/second on target models. The project currently includes an RTX 2070 Super (8GB), an RTX 3090 Ti (24GB), and a 7900 XTX (24GB).

The 2070s is included to see what is possible running the small models, as currently cards with around 8GB VRAM are relatively cheap. My hypothesis is that for a lot of workloads these types of cards are a bargain, and offer a really good opportunity for people to get the benefits of local AI cheaply.

All the components used were bought straight from Amazon, Ebay, or pulled from a drawer. So the only barrier to you doing this right now is approximately £1k, and a few hours of your time to get set up.

I plan on releasing a series of videos build guides and benchmarks as the project evolves. Until then, please fire away with any questions, suggestions, and/or criticisms! 🙂

[Daniel Parke @DanielParke_]

New release for the Hermes Control Hub dashboard, with a lot of major changes🎉

Main changes:
- Control Hub SQLite3 database added
- Mission builder input fields expanded
- Model config page for fine tuning
- Expanded CRUD options
- Gateway streaming chat + history
- UI/UX redesign across most components

Pretty much every system was rebuilt or expanded, so it's easier to check out the repo docs for all changes. Or better yet, just give it a try 🙂

I have also added a user walkthrough guide on how to get started with the dashboard, which is the best place to start for a quick overview of what's available.

I will be monitoring and fixing bugs for the next couple of days, and then I won't be visiting the public version of this again. So feel free to experiment and develop your own version from this template, as there won't be any future changes to wait around for.

As always feedback is very much appreciated, and don't hesitate to reach out if you have any difficulty when trying to use it🙂

https://t.co/kod35fI89o

[Daniel Parke @DanielParke_]

To the nearly 1000 people who have tested my Hermes AI Agent dashboard over the last few weeks, I first want to say a big thank you. The feedback and testing performed has been invaluable so thank you! 🙏

Second to this, is that I plan on releasing a lot of changes tomorrow to the main branch, which should massively improve the dashboard functionality.

I will make every effort to ensure the migration is backwards compatible to support any templates or missions you may have created, but please take some time to backup anything that is really important to you.

I will add more information about the specific changes when the release is ready and on main branch 👍

https://t.co/kod35fHAjQ

[Daniel Parke @DanielParke_]

@badlogicgames You always seem to find the best articles! Thanks for sharing 👍

[Daniel Parke @DanielParke_]

@witcheer Context is killing me when using the Hermes agent, even with a 200k MiniMax window with a 0.6 threshold.

I often find it going over this threshold, which leads to a lot of problems as important context is regularly lost 👎

Better context management needed on my end I think!

[tobi lutke @tobi]

https://t.co/5MJ7u1VHwf

[Mario Zechner @badlogicgames]

recommended reading. i too am concerned about what agents mean for genuine learning. i think what tobi propsoses here is one piece of the puzzle.

[Daniel Parke @DanielParke_]

@Snixtp Amazing resource, thank you for sharing!

[Daniel Parke @DanielParke_]

@KyleHessling1 This looks amazing, can't wait to try it!

[Daniel Parke @DanielParke_]

@7Kiwi The most valuable unit of energy is the one we retain and use ourselves.

The displacement of energy from the grid from domestic solar, increases the available capacity that we could then use to boost manufacturing and other industry.

Less congestion = More connections

[JNS @_devJNS]

me and Claude building an app at 3AM

[Daniel Parke @DanielParke_]

This is really good feedback, thank you!

I had thought about changing the default port to something unused, and making the scripts in general way more robust.

I'll admit, front-end is my forte, so I just took the first option that worked and started tinkering.

Not an excuse though of course, just the reason for the simple first pass! I'm currently remapping out the overall architecture and plan do it properly on the next run. I've a list that keeps growing of things to be refined, so I'll be busy with this project for a while I think 😂

[Daniel Parke @DanielParke_]

Really glad you like it! ♥️

No, I started on it a few days before I found out about the built in one with the same name.

Which is why I changed the repo name to "Control Hub" as well, just to try and avoid confusion 🙂

If you check the commits you'll see that I build about 90% of it with my Hermes Agent (Bob), and then I manually refactored it all to work out the kinks. So we have Bob to thank for the design, I'm just a glorified QA at this stage 😂

This only works with Hermes at the moment, but I plan to add an MCP layer when I get a chance.

I'm in no rush ATM just as I'm having a blast with Hermes, but it would be nice to allow other agents to be controlled from it.