Aayush Panchal · Software Engineer
I build intelligent systems.Now I study how they know.
Currently thinking about: why models trust stale memory over fresh context; what happens when retrieval and training disagree; fitting an 8B-parameter model onto a single GPU; signal, noise and risk-adjusted returns; where machine learning goes next
01 / Background
About
I'm a software engineer who ended up in AI from the backend side: Java, Spring Boot, databases, cloud. That turned out to be a good place to start, because most of the hard problems in production LLM systems aren't the model. They're the plumbing around it: what gets retrieved, how it's stored, and what happens when the model and the documents disagree.
At Learn2AI I work on Cortai, an LMS platform, where I've built a RAG pipeline and fine-tuned open models for tutoring. That work raised a question I can't stop thinking about: when a model's built-in knowledge conflicts with what you hand it, which one wins, and can we control that? That question is what I plan to pursue in a PhD.
I hold an M.S. in Computer Software Engineering from RIT (Dec 2024). Outside of work I write about new models and fine-tuning, and I'm curious about quantitative finance.
Role
- Software Engineer, Learn2AI
Education
- M.S. Computer Software Engineering, RIT (Dec 2024)
Focus
- LLM fine-tuning, RAG, knowledge integration
Based in
- New Jersey
Also curious about
Simulated
Quantitative finance
1,143.71
02 / Career
Experience
Dec 2025 – Present
Learn2AI
Software Engineer
- Building a structured AI tutoring system and RAG agent for Cortai, an LMS platform
- Designed a retrieval pipeline with LangChain and pgvector
- Fine-tuned Llama 3.1 8B with QLoRA for the tutoring system
Community Dreams Foundation
C Speed LLC
NJ Department of Community Affairs
Tata Consultancy Services
03 / Build log
Projects
The Llama fine-tune at a glance
0B
Parameters, Llama 3.1
0 bit
QLoRA quantization
0 GPU
Hardware used
Cortai RAG Tutor
Retrieval-augmented tutoring for an LMS, so answers come from course material instead of the model's guesses.
Internal — writeup available
Llama 3.1 8B QLoRA Fine-Tune
Adapted an 8B open model to a specific task on a single GPU using 4-bit quantized LoRA.
FairLearn
RIT research with Prof. Zhe Yu on fairness and privacy in machine learning.
LogBERT
BERT-based modeling of system logs.
04 / Open questions
Research interests
LLM fine-tuning and knowledge integration
Language models carry two kinds of knowledge: what they absorbed in training and what we retrieve for them at runtime. When the two disagree, models behave unpredictably: sometimes they trust stale memory, sometimes they get misled by bad context. I want to understand and control that trade-off.
Current focus
Knowledge conflicts in RAG systems, using a real problem from production tutoring as the starting point.
Interactive · illustrative
Which one wins?
What is the capital of Australia?
Parametric memory
Learned in training
“Canberra”
Weight 70%
Retrieved context
A stale document
“Sydney”
Weight 30%
Model answers: “Canberra”
A toy example, not a measurement. Real models don't expose a dial like this, which is why the question is worth studying.
05 / Notes
Writing
I write regular, hands-on breakdowns of new models and fine-tuning techniques on LinkedIn and Medium, with diagrams, references, and practical takeaways.