What we've done and learned

TalkToDART(T2D)

Dylan is now finishing up the RAG Engine for T2D. He's done a fantastic job so far, and I really appreciate his discipline. Pierre has been providing valuable insights from the customer's point of view. As a result, our product surpasses other agents in performance and user experience.

Below is the actual comparison of results. The key point is the table presentation of answers. Our product alone aligned with the intent, showing precise data from DART — the Korean equivalent of the SEC's EDGAR. Additionally, T2D analyzes the data to provide a starting point for your own analysis, which is truly remarkable.

You might wonder why we're significantly better than these big tech companies. The answer is their lack of capability and concern in this area.

We view the LLM landscape similarly to the computer industry, where foundation models (e.g., GPT-4, Mixtral, Claude Opus) are akin to semiconductors. Providers like Google and OpenAI, designing models from dataset selection to training methodologies and tuning, resemble fabless companies in a market characterized by a winner-takes-all behavior. Cloud services like Azure and AWS, where the models are trained, act as foundries similar to TSMC or Samsung Electronics.

This perspective highlights that existing companies are focused on preserving their territory. Intel and NVIDIA did not invent the personal computer, nor did TSMC or Samsung create computers. They are concentrated on their core business, which is why we have developed a module superior to other agents.

Advice

Creating a new, unique, and compelling pitch deck that instantly captures the audience's attention is always challenging, regardless of how many presentations you've made. Criticism comes easier when the work is not your own. Hence, I sought advice.

I consulted with a friend who has experience in management consulting and private equity. My friend described the deck as terrible (a sentiment I share), saying it resembled a product manual. Besides that, I was advised on strategy for pitching Pado, with two options in mind:

• Start with ASQ: Focus on what we aim to create, then demonstrate our progress toward this goal.
• Start with TalkToDART: Highlight our ongoing projects, then illustrate our expansion plans.

I published the current version of the deck based on the latter, though my friend believed the first approach might be more effective, recalling our past discussions on LLM applications' challenges and their unique implementation in ASQ.

(Recap) Five pillars of ASQ

The five essential requirements are:

1. Goal alignment: Your intentions must align perfectly with the system, requiring minimal conversational steps.
2. Global task decomposition & management: Once your goal is specified, it needs to be broken down into a task chain in the form of a DAG.
3. (MOST CRITICAL) Local task management & execution: Assign each task to a "module" focused on performing complex functions.
4. Easy revision: You should be able to provide feedback on the system's response to refine the results.
5. Slick interface: The entire process should be encapsulated within a simple and intuitive interface.

This framework has been central to our approach since its establishment in October 2023. We have explored various ideas for achieving these goals, reading papers, testing projects, and developing prototypes. We realized the necessity of possessing modules with exceptional performance. Initially, the lack of useful tools, especially for search functionalities, was apparent. Now, we believe tools like Perplexity could be beneficial.

Pado

We are preparing to open our pre-seed round and will incorporate as a Delaware C-corp, as many of you know. We are seeking over 125K USD for a six-month runway to reach our milestones, which includes hiring new engineers.