This is a full stack application for legal domain inspired by "HARVEY" or "LEGALFLY". I am trying to build this product to understand for building a domain which will only care about the output of the application, so understanding another domain of people will be an interesting approach for me, and then how can i use deep tech and innovate on the backend part on how the reasoning and functionality is working would be interesting. Designing my own architecture for queries, caching, and how i use the agentic workflows will be interesting to see.
This will help me understand what a good UX is, and also if you want to innovate on the backend how you do that and building a simple product can be so interesting.
If my agentic workflow will be better and optimal for a particular use-case will be interesting. The purpose is that i will gain and acquire knowledge for solving a problem that i will have no idea about and how i approach to it, what new i've learned which doesn't exist out there will also be interesting.
I intend to implement things in the backend with as much control as I can, because the frontend element parts will just be taken care of.
Legal Works require a load of paperwork. There are paralegals and there are lawyers, their tasks can be broken down into 3 subtasks, "ANALYSE", "RESEARCH" and "DRAFT". And every line on the contracts or in the legal domains can be cited.
This is the insight and the thesis on which Agent Binod is being build. It will help you analyse contracts/documents, now the research is all about how good of a search engine do you have for retrieval and drafting is how cited and backed each of your line in contract.
I'll focus on the 3 things for building agentic workflows : how good my agent can reason how good my agent can retrieve how good my agent can draft
but there is also one thing which is important which is how good the ingestion of information process is.
Now, we break it down into simpler step by step approach.
Documents and information will be ingested and will be multimodal, like handwritten notes, long contract papers, company financials, party information, can have hearing recordings or personal lawyer can have settlement evidences.
This will be approached by chain of thoughts and understanding every bit of information and making how relevant one thing is to another, as the information grows on a subject how do we keep the memory and not provide wrong information.
There is an interesting approach which suggests that you can make bigger pretrained models fine tune on smaller amount of long chain data, as is if we can give good enough data to how reasoning is done we can get great results.
We answer questions, some internally to mimic reason and some to help the person using it reason .
Whatever reason or node we have, we will back it by articles, or laws or cases, the good thing about legal domain is that everything needs to be documented, and every line is relevant.
This step is where, how we organise and write down things backed by each statement and reasoning will help us, this is the goal for someone using it, here we are a mere tool. We don't make what is right or what is wrong we have set of rules and understanding and citations, we use them and break down every bit of necessary information in the way it should be.
There are templates for each type of document draft, the more people will use the more we'll learn.
We have to use the approach of SOFTWARE 2.0, where we have to write code, as it grows it get's better. The more code write itselves from what people using it are telling the better it is.