Hey everyone,
I'm currently building a startup that's essentially a specialized Notion/Overleaf for university students writing their final theses.
My background: I'm a Data Science graduate currently working in Marketing Ops. I know my way around data and scripting, but I'm not a hardcore ML Engineer.
The Product: We use OCR to scrape university formatting guidelines (margins, fonts, required sections). The user writes in our Notion-like editor, and we ensure their document follows both the strict university rules AND specific custom rules set by their professor (e.g., "Always put the URL when citing", "Use passive voice", or "Expand acronyms only the first time").
The Problem: I'm struggling with how to architect the AI Assistant part. I don't want a Grammarly-style auto-correct where the user loses control. I want a proactive assistant that highlights text, explains why it violates the university/professor rubric, and suggests a change that the user must approve or reject.
I’ve been looking at YC companies like Frizzle (AI grading for tutors) as inspiration. My assumption is that these kinds of startups aren't training bespoke ML models from scratch. They are probably staying "on the surface," acting as wrappers around LLMs combined with some orchestration framework.
Since I’ve been out of the ML and AI loop for a bit, I wanted to ask the community:
- How are modern AI startups actually approaching this? Are they doing deep ML (fine-tuning, custom models) or just orchestrating LLM APIs and tweaking parameters/prompts?
- Do you have any suggestions, mental models, or framework recommendations on how to approach and build this AI layer?
Any advice would be hugely appreciated! :))