Category: Daily Report

Yesterday took a good step forward on a stealth AI project I’ve been involved with since January. It’s operating in a crowded market, but we’ve recently identified a better way to pivot. Will be able to announce it soon, once final alpha tests have been done.

Also this week we have potentially found a new addition to our management team – a proven AI lead generation specialist. But more on that another time.

Today I began taking another look at N8N, I’d already tried a first time and was impressed. I was able to quickly create a workflow that read a webpage, summarised it, and then pumped it to Slack as a message.

Honestly, these automation tools WILL be part of every single company going forward. They will replace a lot of internal software platforms.

My first aim is to go from a WordPress post to publish on:

• X
• LinkedIn
• Facebook
• Instagram

There is already a template setup, which is what’s so powerful about this N8N stuff – most of what you want to automate has already been done.

• Installing N8N with Docker is super straightforward
• Look at the automation templates: https://n8n.io/workflows/
• For local use it’s super easy to create a webhook tunnel

https://n8n.io/workflows/3086-publish-wordpress-posts-to-social-media-x-facebook-linkedin-instagram-with-ai

Today’s video analysis is of the guy who made Windsurf – which I haven’t used yet, but am going to give a proper try. Might prefer it to Cursor!

Main takeaway is:

• Sales team combines AE and forward deployed technicals

In other news, this is a real time object identification locally that I want to try out

https://github.com/ngxson/smolvlm-realtime-webcam

For me Structured Outputs are essential when talking to an LLM. In most cases, an application needs to use data in a certain way so you want OpenAI to send back data in a specific structure, rather than having to format yourself.

Structured Outputs (SOs) ensure an OpenAI response is structured in the way you would like.

Prior to SOs you would use a specific prompt to get the response in the right structure.

Parsing a Rental Property Webpage

Let’s set a schema that will take a blurb from a Cottage Rental website and output an object …

import { NextResponse } from 'next/server';
import OpenAI from 'openai';
import { zodTextFormat } from "openai/helpers/zod";
import { z } from "zod";

export async function POST(request) {
  try {
    const client = new OpenAI({
      apiKey: process.env.OPENAI_API_KEY,
      organization: process.env.OPENAI_ORG_ID,
    });

    const requestData = await request.json();
    const { model, input } = requestData;

    const PropertyListing = z.object({
      cottageName: z.string(),
      numberOfGuests: z.number(),
      numberOfBedrooms: z.number(),
      numberOfBathrooms: z.number(),
      allowsPets: z.boolean(),
      keyFeatures: z.array(z.string()),
      marketingCopy: z.string()
    });

    const response = await client.responses.parse({
      model: model || "gpt-4o-2024-08-06",
      input: input,
      text: {
        format: zodTextFormat(PropertyListing, "property"),
      },
    });

    return NextResponse.json({ 
      success: true, 
      event: response.output_parsed 
    });
  } catch (error) {
    console.error('OpenAI API Error:', error);
    return NextResponse.json(
      { success: false, error: error.message },
      { status: 500 }
    );
  }
} 

I can call this API crudely with something like:

    try {
      const response = await fetch('/api/openai/parse-cottage', {
        method: 'POST',
        headers: {
          'Content-Type': 'application/json',
        },
        body: JSON.stringify({
          model: "gpt-4o-2024-08-06",
          input: [
            { role: "system", content: "Extract the cottage information." },
            {
              role: "user",
              content: `webpage html here`
            },
          ],
        }),
      });

      const data = await response.json();
      
      if (!data.success) {
        throw new Error(data.error || 'Failed to parse cottage');
      }
      
      setEvent(data.event);
    } catch (err) {
      console.error('Error parsing event:', err);
      setError(err.message || 'Failed to parse cottage');
    } finally {
      setLoading(false);
    }
  };

The response I get is

{
  "cottageName": "Point Break",
  "numberOfGuests": 8,
  "numberOfBedrooms": 4,
  "numberOfBathrooms": 3,
  "allowsPets": true,
  "keyFeatures": [
    "Beachside location",
    "Contemporary interiors",
    "Part of The Dunes development",
    "Traffic-free location",
    "Open plan living",
    "Floor to ceiling windows",
    "Balcony with ocean views"
  ],
  "marketingCopy": "A beautifully finished beachside house over three floors with contemporary interiors to the highest standard. Part of The Dunes development located directly in front of the beach. Tucked away in an almost traffic-free location. Open plan living, dining and kitchen space on the 2nd floor, taking advantage of the breath-taking views over the golden sand dunes, beach and Atlantic Ocean. The spacious living area comprises of two large sofas, TV, DVD player, coffee table, floor to ceiling windows, patio doors opening onto the balcony."
}

So, what that’s done is convert some HTML into a JSON object that summarises the HTML. Useful!

For today thought I would just add my first Sora video.

The generative industry is moving so fast; with images, video, music, speech, written word all now improving at a rapid place.

Many people are not seeing it.

Today I was able to build a web tool that may not be great for production but is really quite good for the debugging and investigation of an API. The ‘AI’ was able to understand some of my complex but no quite so explanatory prompts in a way that blew me away and was able to produce some really valuable work.

I don’t fully know whether programming is going to die as a profession. I think *eventually* yes, it probably will be abstracted away much like assembly language has been. All the patterns will be fully understood, trained, tweaked and they will be able to replicate huge swaths of code that actually create a great product.

When you do get into complex systems though – whilst I’m no silicon valley person – it does seem that there are limitations on it, and it must have an experienced human (developer) guiding it to get it to the last 5% of properly working.

That’s just programming. The rest of the generative industry is equally as mind blowing.

Are people really seeing what’s going on?

And do they understand the magnitude?

And are they embracing?

Work Diary

Laravel Herd & Filament

Whilst I have been prototyping with Vanilla, I checked out Filament last night on Laracasts:

Filament Forms

The Form library alone is complete enough to warrant using Filament. It has so many small quality of life features that add up. From memory here’s a few that convinced me:

• Really nice prefix and suffix options
• Really nice icon and description and helper text integrations
• Excellent search integration into long selects in one line of code
• A really easy wizard workflow setup
• Excellent integration of new entity creation when a user needs to be able to add a new option to a select list
• Easy linking between SELECTS

This is of course on top of it working solidly as you would expect.

So, I’m flipping my prototype to use Filament … which is an excellent basis going forward.

Filament File Uploads

In one line, its got a drag and drop file upload, with image preview built in. Uses filecon JS

Today, I was looking at making a final decision on whether to use a React or Vue frontend. My prototype has been coming along, but I had just been using plain PHP blade templates with Vanilla Javascript for quickness.

I think web development has got totally overcomplicated, and that you can go a long way with vanilla stuff. I would never start a PHP project without Laravel however, since I know it has virtually everything I need to build whatever I want.

Since LLMs have come about, there’s a lot to be said for just writing functionality in javascript; but ultimately you do want to benefit from a frontend framework. State management, DOM manipulation and an event bus are reasons why I want to use them. And of course, components. Was a shame the official WebComponent standard is so poor.

React Or Vue? Or Livewire?

The age old debate for a developer. It’s annoying quite frankly at this point to continually decide what to use. I prefer the simplicity of Vue syntax, but I often use graduates from a bootcamp that teaches React, and I have bought quite a few premium templates that React also. To be honest though, with LLMs it’s so easy to switch between.

Then there’s LiveWire. Each time I’ve tried to get into LiveWire I’ve been put off by it. But I had the same resistance when I got into Flutter and once I persevered I loved it.

In order to decide on the choice, I looked into Filament.

Filament

Filament is TALL stack (Tailwind, Alpine, Laravel, Livewire) … and i’ve used previous versions being fairly impressed. This would be a bit of a few steps back in order to take a leap forward, but taking a look at the excellent Laracasts for it, I’m starting to see the benefits.

• The resource creation gives the CRUD UI upfront … this is slightly different from the CRUD UI generation that I have been working on… since my ones are more ‘on-the-fly’ than ones that get defined in the yaml files. Still the resource creation results in a full on CRUD interface with sortable datatables out of the box.
• The form library is pretty much flawless. No more messing about with forms.
• Select options can be configured with enums and gets all enforced really nicely.
• Multi-tenant stuff out of the box
• Can hook up with Laravel Stripe for subscriptions

Laravel Herd

• I’ve been a fan of Laravel Sail for a while, but that’s only because it abstracted complication away from me. Whilst I get tired of learning new things, that’s no reason to not embrace something that works even better. Enter Laravel Herd. Some benefits:
  • I don’t have to use Docker Desktop. Which decides to hog a huge amount of system memory. You can change this in settings, but overall Herd is a lot faster.
  • For some reason, my local composer would not install the laravel CLI tool to my local machine. So I had to resort to installing Laravel via some shell scripts. Herd got it installed immediately
  • Herd includes TinkerWell, some debugging tools, and general system configuration

Consultancy Fees

I’m getting tired of being asked to ‘look into stuff’ and expecting it to be done for free. Need to have a consultancy page hooked up to a simple credit card thing.

I’ve been using my very basic prototype to store data of ‘things’ in my life – I use LLMs to setup this data; and then I have a UI that allows me to interact with the data.

One of the things I’ve been doing with it is keeping track of projects, ideas and tasks … and the problem I started to face was I didn’t like the UI I had used as the index for these data items.

So last week I’d spotted a nice slider effect that I liked; and today I decided to put time into my own project for once again; and just implement that.

So now I still have a CRUD interface for my data, but I’ve also integrated the beginnings of a nice swiper interface so that I can swipe through each ‘thing’ one at a time; consider it one at a time; and then move onto the next.

I made an animated gif to briefly showcase it. Since I’m running out of time today, I’ll leave it here … but this is currently cycling through my blueprints … it’s a good start and going to be a much more enjoyable way of cycling through the things going on.

And of course eventually the images will be generated according to the content.

More updates tommorow.

It’s been a few days since I posted last. I’m going to let myself off, but resolve to do better in the future – since I do want to keep posting on a daily basis but I am aware I want to make the posts more valuable; so I feel they represent value that can then go onto LinkedIn, and the other social networks.

Things are definitely building in my mind; and projects are coming along. However, still a long way from anything concrete taking off.

For the moment would just like to talk about MCP, because it really is a very significant milestone in the AI world. (Image context from OpenAI was another one recently).

MCP

Model Context Protocol was announced in Nov 2024 by Anthropic.

This is a big step forward since it has standardised how language models can/should interact with external tools.

In order for LLMs to be further useful, they need to be able to action things.

I’m guessing that once Anthropic had integrated a couple of tools with their LLMs they realised it would be better to have a standard way of doing that … and from that they came up with the MCP architecture.

At first understanding, giving a language model a standard way of interfacing with the outside world is a bit of a game-changer and there are already a ton of integrations that we can use straight away. Things are happening really fast.

Core Architecture

Essentially, you have an application (called the Host), and this has a Client inside it.

The Client maintains connection to the Servers which have access to the tools you want the LLMs to connect to.

Protocol Layer

It’s fairly simple … you have:

• Requests
• Notifications
• Results

You have functions that:

• Handle incoming requests
• Send requests and await response
• Handle income notification
• Send a notification

Transport Layer

All transport layers use JSONRPC

JSON-RPC Request Object

• jsonrpc : always going to be “2.0”
• method
• params (optional)
• id

Without an ID, it is considered a notification that doesnt expect a response. In fact, servers MUST NOT reply to a notification.

JSON-RPC Response Object

• jsonrpc (version string)
• result
• error
• id

To be clear, MCP is a standard way – a proposed specification that multiple parties can agree on – for how language models will interact with outside tools.

Protocols are vital in tech. We have TCP-IP and email, which you use all the time. Without that agreement, we could have had an incredibly fractured internet.

When companies and developers can agree to do things in a certain way, it makes it easier to make systems.

Using MCP I assume is very much like working with interfaces. If you code an LLM up to work with your own tools and then for whatever reason you decide to switch LLMs … then using a protocol would mean there’s no lost work – because the new one will use exactly the same interface as the current one.

Reference

4 Request object

A rpc call is represented by sending a Request object to a Server. The Request object has the following members:jsonrpcA String specifying the version of the JSON-RPC protocol. MUST be exactly “2.0”.methodA String containing the name of the method to be invoked. Method names that begin with the word rpc followed by a period character (U+002E or ASCII 46) are reserved for rpc-internal methods and extensions and MUST NOT be used for anything else.paramsA Structured value that holds the parameter values to be used during the invocation of the method. This member MAY be omitted.idAn identifier established by the Client that MUST contain a String, Number, or NULL value if included. If it is not included it is assumed to be a notification. The value SHOULD normally not be Null [1] and Numbers SHOULD NOT contain fractional parts [2]

The Server MUST reply with the same value in the Response object if included. This member is used to correlate the context between the two objects.

[1] The use of Null as a value for the id member in a Request object is discouraged, because this specification uses a value of Null for Responses with an unknown id. Also, because JSON-RPC 1.0 uses an id value of Null for Notifications this could cause confusion in handling.

[2] Fractional parts may be problematic, since many decimal fractions cannot be represented exactly as binary fractions.

Strong results come from doing the right thing many times over.

Consistency. Discipline.

I’ve got multiple projects going on at the moment and they can easily fill up my days entirely, without time for the things that will necessarily continue to build the business. But I am dropping things that will actually continue to build the business.

These are a good additional starting point to my existing habits:

• Grant and funding research & application
• Event, conference and networking research
• Email marketing pipeline
• Website marketing
• Social media marketing

It’s a bit like programming yourself for succeeding; you decide what actions the business would benefit from; or what areas need to be attended to consistently.

These become the KPIs of your business.

I think the way the world is right now, people need to feel like there is a new wave of ‘something good’ … and whilst AI is certainly going to take many jobs, it opens up a whole world of possibility for those who are prepared to learn about it, work hard and creatively use it. So there has been this huge positive wave of energy (certainly amped up with money) toward innovation. It may well be that LLMs have limitations, and that we are witnessing a bubble… but people are now TRYING new things. Technical and non-technical people are finding they can do a whole lot more using AI. What I mean is, a lot of crazy ideas are being unlocked. Pandora’s box has been opened.

True Personal Assistants

For me, it gives me the opportunity to build a personal assistant of the magnitude that i’ve wanted to few decades. I think the science fiction movies and video games influenced me in wanting to build these personal assistants.

Examples are

JARVIS Assistant From Iron Man Movies

J.A.R.V.I.S., which stands for “Just A Rather Very Intelligent System,” is Tony Stark’s natural-language user interface computer system, named after Edwin Jarvis, the butler who worked for Howard Stark and the Stark household.18 Initially, J.A.R.V.I.S. was developed as a simple AI assistant to control Stark’s Iron Man suit, but it evolved into a powerful AI capable of managing various tasks and assisting Stark in his personal and superhero life.51 J.A.R.V.I.S. uses advanced natural language processing and communication skills to understand and respond to Stark’s commands, making it more than just an AI tool—it’s a trusted companion.

So, the ‘trusted companion’ thing here is the key.

Cortana from Halo

As an artificial construct, Cortana has no physical form or being. Cortana speaks with a smooth female voice, and projects a holographic image of herself as a woman. Cortana is said to resemble Halsey, with a similar attitude “unchecked by military and social protocol”. In Halo: The Fall of Reach, Cortana is described as slender, with close-cropped hair and a skin hue that varies from navy blue to lavender, depending on her mood.^[6]: 216 Numbers and symbols flash across her form when she is thinking.^[9] Halsey sees Cortana as a teenage version of herself: smarter than her parents, always “talking, learning, and eager to share her knowledge”.^[6]: 218 Cortana is described as having a sardonic sense of humor^[10] and often cracks jokes or wryly comments, even during combat.^[6]: 217

There’s a few more but that’s enough for now

So, the key components are:

• LLMs (Language Models)
• Decision Trees
• GenerativeUI
• Automation flows
• Data Mining & Analysis
• Context and Object Oriented Memory
• MCP

In the future I imagine, the internet has pretty much been abstracted from us. Which isn’t a great thought for most of us, but it’s kind of a natural progression. Things change, for instance it was naive of me to think that I could have another 20 years of making money doing the same thing until I retire i.e programming.

At school, I had a maths teacher who used to program in assembly. Talk about talent… in a few decades will we have any humans that know how to write in assembly? I always wondered why he didn’t still program, and it was probably because he didn’t move on.

In the same way, those of us who understand HTML/CSS/JS on a deep level … these skills are fast becoming abstracted away with vibe coding apps. The vibe coding apps that combine with Supabase are going to get better and better; but will the prediction machine ever get so good that it can really understand programming so that it makes it super clean (probably yes, as our prompt engineering gets better to help it). Eventually, the personal assistants will replace any requirements that vibe coding creates now.

The next big waves/industries to come along will be more widespread autonomous machines (drones, robots, cars), surveillance and wide scale sentiment monitoring; then beyond it some sort of biohacking industry specifically with longevity at the core; video games will have a resurgence in originality; so I do hope that something will replace the employment drop-off that will shortly happen.

That’s enough thoughts for today.

Work Updates:

• I continue to work on a side project to bring in a small amount of money but enough to cover personal costs. It’s a fun, high potential project, and made me realise I do like working on IT dev projects, and am quickly improving at the management of them.
• The DXP project is slowly getting to the beta launch was is great. Once that’s hit, I will share more details.
• My own platform is coming along. Shown it to a few people and I know these people wouldn’t sugarcoat stuff, it’s super days but reasonably positive. Just got to keep going.
• The ideas that would drive an AI Agency are coming together; but still a little way off from doing that
• Continually trying to keep up with the industry
• More and more aware of the necessity of marketing for everything