Arka Works Ltd

"I need to compare product specs, but the AI isn’t reading the data correctly..." To achieve the highest levels of accuracy with LLMs, we must look at what we’re feeding them. LLMs love plain text, but they do even better with Markdown. Markdown is like plain text, but it includes the ability to communicate basic formatting. With most consumer apps we have all kinds of issues with PDFs when they are loaded as attachments. Often they aren’t read in full, or there are formatting issues with data, tables, diagrams. We are in the final stages of testing before integrating an outstanding new model called “Mistral OCR” into OmniChat.uk. Mistral OCR was launched in the last fortnight and we are planning to include it as an add-on for Business level subscribers as a solution for long document markdown conversion, something I'm asked for often... This tool will massively improve the accuracy and performance that we’ll be able to get from PDF attachment inputs because they will be scanned in full and converted into Markdown BEFORE they get fed to the LLM. OCR stands for “Optical Character Recognition”, or the ability for the model to detect and extract info from all manner of graphical input material and turn them into a digestible format for any LLM.  Mistral OCR performs at 94.89% accuracy and with the ability to process up to 2,000 pages per minute, it beats current offerings from Google and Microsoft, particularly for handling very complex documents that as professionals we need to be able to work with. TDLR, don't use standard attachments... there is a better way.

I got to sit down with Chris Simmons this week to talk about AI and Architecture and mostly to reflect on the many incremental career steps that eventually led to setting up Arka Works Ltd and now OmniChat.uk. It's quite rare that we take a moment to look back and post-rationalise all of our career decisions and what they all meant ... it was a bit like therapy! It was my first time meeting Chris in person, he's very charming and he IS definitely taller in person 😊

Often, the best discussions happen in person and don’t involve structured agendas or take place over teams meetings. I travelled to London today to meet with Keir Regan-Alexander of Arka Works Ltd as part of our collaboration at Henry J Lyons Architects. Keir and his business partner Stephen had been working on some exciting new ideas in the background. He was able to show me enough to give me that lightbulb moment without telling all just yet. We did a deep dive into some prompting workflows in OmniChat.uk and tested his recent discoveries of model accuracies with a prompt on the limits at 300k tokens - jumping from Gemini flash 2.0 for the initial prompt, to GPT o3 mini’s reasoning model to improve the result and then back to sonnet 3.7 to apply some structure and check the accuracy, exploiting each of the model’s potentials, crashing chrome, starting over again and grinding out a result within a short time period which, we could benchmark against days worth of work recently completed with a manual approach. If this wasn’t enough we talked about the potential of agentic workflows to provide the rocket fuel for some of the existing prompts we’ve created which have already demonstrated high-impact results. We also discussed good coffee and shared sourdough tips, for fear of the day becoming too AI-dominant. #arkaworks #henryjlyons #hjl #ai #gpt #claude #gemini #architecture #aiarchitecture #omnichat

🚀 Exploring AI in Architecture – Plus free Pizza & Beers! 🍕🍻 Curious about how artificial intelligence is transforming the architecture, and the engineering, and construction (AEC) industry? Don’t miss the Q1-2025 Design Technology User Group (DTUG) meetup in London, taking place on Tuesday, March 18th, generously hosted by KPF. We’re excited to present an incredible line-up of speakers who will share insights on the practical applications of AI in AEC, including: 1️⃣ Keir Regan-Alexander Co-Founder @ OmniChat.uk & Principal @ Arka Works Ltd ⚡'LLMs and Good Prompt Craft'. 2️⃣ Allister Lewis Founder and Director @ Automated Data Driven Design ⚡'AI and Generative Design in Practice'. 3️⃣ Giuseppe Bono Head of Design Computation @ tp bennett ⚡'When AI Meets Architectural Practice: Between Strategy and Practical Implementations' 🎟️ Tickets are always free, but limited, so don’t wait! Find the link to secure yours in the comments. For more information, feel free to reach out to me or connect with the DTUG organisers: Daniel Giddins @ Symetri Europe Martin Wilding @ KPF Mike Turpin @ Connected Places Catapult & Innovating Futures Ltd Caoimhe Loftus @ Autodesk Forma Wayne Patrick Dalton @ Pascall+Watson We look forward to seeing you there! ✨ #ai #architecture #aec

My friends tell me they have a real love/hate relationship with LinkedIn. Some find it cringe-worthy, others feel obliged to participate for work. Many log on, feel overwhelmed by the achievements and exclamations on their feed, and feel a sense of panic. About three years ago, I made the deliberate choice to quit Twitter and remove Instagram from my phone for two main reasons: 1. I was addicted, and it was a very poor use of my time. 2. Anonymous posting often leads to appalling behaviour with no accountability. So, I redirected my focus to LinkedIn - a platform where you post under your real name, with your professional CV attached, and where dialogue tends to be much more productive. Since then, I've experienced real benefits from becoming an active contributor rather than remaining a passive observer. Numerous opportunities have arisen through the connections made here. As Paul Graham states, "Writing IS thinking" and every post or newsletter I send out (sorry they're always so lengthy!) forces me to clarify my thoughts and the strategy and actions I should take next in a public forum. Posting does involve sticking your neck out a bit. You worry someone might disagree, ridicule you, or worse - tumbleweed. But if you can post without an overwhelming fear of bad reactions, that's some form of superpower. This week, I crossed the 10,000 follower mark. By social media standards, that is modest - but these are strong, productive connections and I'm proud of them for a few reasons. There are plenty of influencers here advocating for formulas to achieve "massive growth and engagement gainz" - hacks for specific posting methods, comments and writing styles. Perhaps that approach works for building a large network of loose connections. But I'm not convinced it's worth it. Posting when you feel the urge, writing in your genuine voice (not using AI!), and sharing authentic insights seems to lead to real-world outcomes derived from stronger ties, that are very rewarding. When I meet people in practice for the first time, they often already know about my work and mention something I posted recently. At events, people now come up to chat about shared interests and I love that. These connections have also directly resulted in the majority of my enquiries since I launched. Almost any professional update can provide value to someone else if considered thoughtfully. Before I post, I always ask myself: "Does this offer something useful to another person in a similar position?" If the answer is yes, then it's probably worth a post. So, if you've made it to the end, thank you again for following! (And thanks to Finn and Theo for their artwork at the breakfast table today)

Efficiency, workflow, new business, creativity, and sustainability are at the forefront of minds in the industry. But how can we use #AI and new digital tools to support and achieve these ambitions in practice? Join us on Thursday 20th March to learn about the opportunities, challenges, and real-world applications of AI and new software tools in shaping the future of architecture and construction. Register for free here: https://lnkd.in/e7y5x-_7 Speakers: - Keir Regan-Alexander, Principal, Arka Works Ltd - Luka Stefanovic, Senior Architecture Industry Specialist at Vectorworks - Lilian Ho, Senior BIM Manager & Digital Lead, AECOM and Program Director at ZIGURAT Institute of Technology - Bret Tushaus, Vice President, Product Management, Deltek - Nasrine Tomasi, Head of Artificial Intelligence, Mott MacDonald

I'm planning to use this speaking slot just to talk about LLMs and good prompt craft - come along if you'd like to hear more :)

What is "Unabridged Prompting"? In short it means dropping all your input words and data IN FULL into a very large context window and avoiding using attachments. With the right setup you can drop ~750,000 words into a single thread, opening up a load of use cases in practice (Gemini 2.0 Flash = 1m tokens). When you load attachments into consumer apps they get "RAG processed". It's a bit technical, but this means the model will retreive the 'most relevant' and similar chunks of text from the documents ... but when you're dealing with complex and nuanced material in professional settings - you need the model to reason across the whole body of text in its entirety. When you use unabridged prompting you will see immediate improvements in: - Accuracy of outputs - Retrieval of facts and ability to provide citation across the whole doc - Ability to reason and synthesize information more effectively To access the full model context windows you need to be using the API versions. Most apps are much smaller (32k window). For the absolute best results, you want to prompt with full length text, ideally formatted in Markdown to preserve headings, subtitles, tables, diagrams - and actually have them understood.

Last week I spent a whole day writing new workflow prompts with a practice and I've started to use a food-based metaphor to explain my process 🥗. To get to the next level using LLMs, you need to shift your thinking and go from a method of 'ad-hoc' conversations prompting over to 'recipe prompting'. When I meet people using language models for work, most are following an unstructured and conversational style. This is all well and good - it's likely to bring lots of value to your every day work, but in order to get to the next level you really need a more structured approach that has a bit of strategy to it. Any task that you are likely to carry out more than once could (and should) be turned into a prompt formula or template that you use again and potentially share with your team. 🗒️ These recipes are called 'Pre Made Prompts' and they can unlock loads of value when you start to think of all the places you can use them across practice. The best metaphor I can think of to describe my process is a cooking one. You have a workflow idea, let's think of that as a salad recipe. It's one that you know works well and always produces a great dish at the end; always made to the right portion size, elegantly presented and delicious. Your great salad recipe has some flexibility built into it - you can change the oil, the seasoning, the toppings, the protein, you can even chop parts into larger or smaller pieces, or you can present in different shaped bowls or plates ... you get the idea. Any part of the recipe that you want to add flexibility to is called a 🥑 "variable", these are the changing ingredients of the recipe. A {{variable}} acts as precise insertion points for new material into your prompt recipe, they get dropped into the mix at the exact moment and order that you specify. Using the right tool and method will allow you to preload all your recipe instructions with one-click and then you simply pipeline the new data to run through them (these are the changing ingredients)... Variable data might be... 🥕 {{Meeting Transcript}} 🥒 {{Architectural Specification}} 🥔 {{Appointment Contract}} 🥝 {{Bid Invitation Docs}} 🫒 {{Decision Notice}} ... {{Literally any standardised piece of data}} My list of ideas for variables is very long - yours will be too. 3 main benefits of this approach: • Consistent, high quality output across all team members • Speed of repitition • Enormous flexibility across a typical practice This is the method we endorse when applying OmniChat.uk in practice. 👉 www.omnichat.uk Happy prompting.

Meetings have always felt like a lot of work, but now they really are. I've become a bit obsessed with taking transcripts in meetings - and I always ask before I turn the transcript on. Using long context window LLMs and prompting with full unabridged length means we can get huge value out of raw meeting transcripts now. We don't need a "meeting minute app" or "copilot to produce a vanilla summary", we can just use the latest and best foundation models and be highly prescriptive about how we arrange the information. After a meeting I produce notes straight from the transcripts (.txt file is best) - they are organised by conversation theme, arranged to define "Areas of consensus" "Areas for further discussion" "Actions and due dates" "Next meeting agenda" For each we can provide full citation to support and evidence all of the notes (because we are using large context window). Then when I issue the minutes, I do a final spot check and I send the transcript along with them - that way everyone has access to the full record. The reason this is different is because when you know an AI can read back through your words exhaustively, you start to speak differently... You begin saying things out loud for clarity, being very specific about actions, responsibilities - making sure to track back and clarify things that aren't clear. All things we should probably do anyway. They represent good meeting craft. The reason this counts is that it's now really worth the effort. The AI is extremely good at picking through the nuance, it will track your findings and it will synthesize everything really beautifully for you. This means we can actually converse with past meetings, ask what a person said about this or that, recall where you landed on a particular issue weeks or months afterwards. Even better though, we can use the transcripts to: - Actually write new technical material, - Draft or redraft whole sections of reports, - Update action trackers (all with formatting), - Create flow diagrams and organograms. ...and many other things besides. If meetings always felt like a lot of work with not much to show for them afterwards I think the whole purpose and culture of meetings is about to change and for the better. The meeting, the long form discussion, the hard won consensus IS the work - you will let the AI write it up for you after. If you want to try these methods I recommend Claude 3.5 Sonnet (200k context window) Gemini 2.0 Flash (1m context window) These work great with a pre made prompt method using multi step instructions on OmniChat.uk Ps thanks to Ayre Chamberlain Gaunt Ltd for this very substantial AI practice meeting last week 🤩