AI Solutions Architect - In pursuit of, and updates on that path.

 Hello!

This post is long, so here is a teaser of sorts:

TL;DR

In this post, I reflect on my deep dive into AI agents, their potential to push the boundaries of artificial intelligence, and my vision for creating a domain-specific, agent-driven ecosystem to automate accounting processes. I explore how agentic architectures could lead us closer to AGI by mimicking the human brain and address persistent LLM limitations like memory gaps through a tiered memory system powered by agents.

---

Highlights

• I’ve been exploring AI agents as the next big step beyond monolithic LLMs.

• Agents offer flexibility and could mimic specialized brain regions for smarter, more reliable AI.

• I propose that agentic, neuromorphic systems may get us to AGI faster than just scaling up LLMs.

• Identified a major limitation: LLMs lack persistent memory, which hinders long-term project work.

• My solution: a 3-tier advanced memory architecture using both large and small agents.

• The aim: build a domain-specific AI ecosystem to fully or near-fully automate my accounting practice.

• Long-term vision: create an accountant-designed, modular system that unites multiple functions into one powerful, intuitive platform.

• Future focus: explore agentic evolution and how AI can move from training to true learning across generations.

Still with me? Here we go!

Been a little while since I have posted. But busy the past month, much of learning and talking with LLM's like ChatGPT, Gemini, and Claude. Its amazing how fast time flies when you're asking questions to the point where its almost like an interrogation. Can't say I'd question a person nearly as much as I do these models. Not because they give so much bad info, but because I want to be relatively certain that the info they are providing isn't an hallucination, and is grounded in actual data, research, etc. 

That said, recently one of those chats was aimed at a deep dive into what is next for AI after agents. I'm sure you may have heard of them by now. They are definitely the hotest topic in my Facebook feed. You can't read about anything related to AI anymore without 'agents' coming up. So naturally, I have been looking into what agents are, and have been curious as to what they can do. Not to mention how far they can carry the technology. The flexibility in their use cases is remarkable to say the least. 

Which has lead me to consider the current state of AI, and how to push the boundaries of it. As it stands, there seems to be two schools of thought. The first sticks with the current evolutionary path of single, or monolithic LLMs - make them bigger, faster, smarter, and more tools/functions. The other involves "agentic systems", which is where my focus has turned and I have found to be the most interesting. 

Before I go on, a small reminder that this is a blog, not a white paper. So I am intentionally being rather causual rather than technical in my writing. More on that stuff later, which you will be able to find on Github when the time comes, as well as other platforms where appropriate. 

With agents often having specialized roles within the system, I started thinking - what if the system was modeled after something familiar - the human brain, for example. It has specilized regions. What if a system were designed to mimic those functions? Or several of those syustems linked together to form the agentic equivalent of a brain? Apparently, I am not the only one who has posed such questions, as research is already underway in pursuit of "neuromorphic" systems that conceptually are aimed at creatinga AI version of the brain using agents. In my humble opinion, this is the path that will get us to AGI, or artificial general intelligence. either faster, or more reliably than a single monolith stuffed full of data. 

Which brings me to some concepts that I have been exploring to eliminate or mitigate issues that I have noticed when using AI - particularly with coding inconsistences, and other particularly long convos that involve project planning and gathering info to move forward with a well-developed roadmap and enough detail in the steps involved to get from A to B. This issue and the solution were noticed and conceived months ago, and there have been advances in the foundational models and their chat interfaces to address their shortcomings, but they are still what I believe to be bandaids at best presently. 

LLMs lack persistant memory. Anyone who has used ChatGPT may be familiar with memory filling up quickly, or that Gemini doesn't recall anything from one convo to another. Claude I have less experience with, and use for very specific purposes that require fine detail that only it seems to provide in certain use cases.

The solution I came up with was posed as 'how do I take a model like chapgpt and give it long-term memory?' or expanded memory beyond what is has now, and is capable to continuing to add to it so it can function as a meaningful assistant and be consistent when discussing project ideas and requesting data-backed advice. RAG and vector databases, as you may be aware, are the most common ways that others have dealth with this shortfall. Recently, during a conversation to try and figure out a more robust and intuitive way to fill the void, the notion of a 3-fold, or tiered approach, was conceived/. Without getting into the specifics (there will be a white paper posted once it has been tested ready for interested parties to put to use and provide feedback), this system, or architecture, creates an advanced memory system using agents, both large and small. There are others who are working on their systems, but my research has shown it to be advanced enough to rival solutions put out by the big names in AI. We will find out if that's true soon. 

Should all go according to plan, this system will form the foundation of a domain-oriented ecosystem of sorts - primarily directed towards automating accounting processes as my motivation as a sole practitioner is to offload what I can so I can better allocate my time. Ultimately, my goal, lofty as it is, is full-automation to near full-automation of my entire practice. I spend way too much time in front of a computer and not nearly enough engaging the people I am trying to help. In pursuit of that goal, many of my projects, which are modular in design, put together will essentially function as my own in-house custom built accounting and tax software (yes, that includes putting it in front of the IRS for testing). Ideally, once deployed, it will perform all the necessary functions to get us through the accounting cycle for each client, or enterprise reliably and intuitively. All the features and functionality that are spread out amongst multiple apps all connected in a modular fashion - possibly and hopefully will be the best you can find because it was conceived and developed by accountant specifically for accounting - something most solutions out there lack. Most developers dont have the domain experience to think of all the little details or features that we wish an app had to make our lives easier. i intend to fill those gaps and eventually market it to others who are like me and just want more of their life back so we can spend it doing the things we love. 

Check out my next post for what may be my most ambitious project yet - agentic evolution and the gradual shift from training to learning - and retaining that knowledge and experience throughout generations.