"Bending LLMs Into Useful Shapes"

Driving question = “how to use GenAI and LLMs truly useful-ly”

I am a recent-law graduate and soon-to-be Trainee Lawyer.

My idea is that we should “bend” LLMs into use-case-ready tools that truly, objectively, unequivocally, and dependably help us produce better work products, quicker.

From my experiments and observations, this “bending” will almost always involve building a layer of Self-designed Parts - Python programs, JSON data models, vibe-coded UIs, etc.

This presentation will display my LLM-based Python experiments: both Manually Coded and Vibe-coded. I intend to walk the audience through:

(i) The 4x tools I have built over the last few years, including (a) a line-by-line walkthrough of the the Python code I wrote to implement my own TF-IDF (lots of Lists and For loops and Dicts) and introducing the packages I used (NLTK, re, numpy, heapq, Pydantic, tkinter), (b) the way I prompted AI Assistants to help me build UI and other functionalities, and use / testing / comparison of multiple different LLMs - Sonnet 3.0 and 3.5, GPT 4o, DeepSeek, QWEN, and (c) a detailed walkthrough of the code I used for Pydantic data model extraction of Personal Information from a variety of uploaded pdfs of fact pattern samples.

(ii) The challenges I came across - e.g. getting the NLTK package to even download and work, how to break my project vision into building blocks, trying to figure out the ways I could built FOR loops to produce the outputs I needed, how to prompt and re-prompt an AI assistant to fix my problems whilst trying to save prompt-tokens/money, figuring out intuitive UI for users, figuring out why my tool works on my machine and not other friends machines, despite being deployed via URL using Replit

(iii) How I solved my problems - actually not that much use of AI (e.g. Perplexity) - I went old school and read blogs that I googled and got led to Medium, KD Nuggets, Analytics Vidhya, GeeksforGeeks, etc. I also learnt tricks on how to “think” and “morale boost” my way out of / over obstacles when I got stuck. Additionally, how I networked and asked people for help was also really rewarding and a huge piece of personal growth that has happened over the last few years of my project journey.

(iv) Some insights on the challenge of getting Data for testing, too - networking around and cajoling / politely asking for data in effective ways is a bit of a process but so important!

• TOOL DEMOS

  • 0) Self-implemented TF-IDF in Python; I draged my .py file into Replit then Prompted it to built a UI for me. Am Attempting to upgrade into an idea called “weighted search”


  • 1) I Hardcoded a Self-designed, Advanced Prompt into Drag-and-drop web app, for document summarising (i.e. “whats in this file, I dont have time to click into it and read it”)


  • 2) Replit-vibe-coded RAG engine for “getting LLM to ‘read’ my books and documents for me” - I can then just ask the app questions about the file/100s of files. Useful for both studying a document, or just understanding what is going on in a hard-drive’s worth of data and files.


  • 3) Python code via Pydantic connected to LLM - used for specific, formatted data extraction from Text Files