Computer Science Undergraduate · Researching NLP & LLM Agents
Hi, I’m Ebaad Raheem.
My research looks at how LLM agents remember accurately over long conversations, and how faithfully models reason under context pressure and quantization.
About
I'm a Computer Science undergraduate at the University of the Punjab, researching how LLM agents can remember accurately over long conversations, and how faithfully models reason under context pressure and quantization. Alongside this research, I build production LLM-agent systems professionally, which has shaped how I think about where these failure modes actually show up in practice.
Research Interests
Toolkit
Publications
Context Collapse in Long-Horizon Agents: Benchmarking Hierarchical Memory against RAG and Summarization
First Author
Advisor: Prof. Ghulam Mustafa · Submitted to the PALM Workshop, NeurIPS 2026
A controlled benchmark of four LLM-agent memory architectures — full-history Baseline, Rolling Summary, RAG, and a Hierarchical (working / episodic / semantic) Memory system — spanning 3,200 conversational runs and 16,000 evaluation points across four professional domains. Hierarchical Memory statistically significantly outperforms the full-history baseline (41.6% vs. 40.6% Factual Retention Rate, p = 0.03) while using 34.8% fewer context tokens, and the study identifies a conversational primacy effect and severe domain-dependent brittleness that persist regardless of context length.
Does Quantization Change Small Language Models' Responses to Misleading Context?
Co-Author
Advisor: Prof. Ghulam Mustafa · Submitted to the LIGHT Workshop (Small Deployment Models), NeurIPS 2026
A matched study of whether post-training quantization (BF16, 8-bit, 4-bit) changes how small instruction-tuned language models respond to misleading in-context cues, producing 3,078 cued chain-of-thought responses across five checkpoints. Using question-clustered GEE models and cluster-bootstrap inference, we find no detectable average effect of quantization precision on unfaithful reasoning, but a robust +8.3 pp sycophancy effect that survives a full-text answer-field audit.
Independent Research
GPT-Urdu — Character-Level Decoder-Only Transformer Trained from Scratch on Urdu Poetry
Self-Directed Research
A decoder-only Transformer language model implemented entirely from scratch in PyTorch — causal multi-head self-attention, pre-norm residual connections, learned positional embeddings, and the full autoregressive training loop — rather than fine-tuning an existing checkpoint. Trained a compact 3.2M-parameter, 6-layer / 8-head model at the character level (59-character Urdu vocabulary, no subword tokenizer) on the Urdu Ghazals Rekhta poetry corpus.
Held-out compression reaches ~2.43 bits/character versus a 5.88-bit random baseline after 10,000 training steps, with nearly identical validation and test loss — evidence the model learned real structure in Urdu script rather than memorizing the training data. Motivated by Urdu's severe under-representation in mainstream NLP research and its distinct right-to-left Perso-Arabic script.
Selected Engineering Work
Production systems built alongside my research, including agentic AI integrations and cloud-native backends.
SkillMorph
An online course marketplace (create, sell, and buy courses) with Stripe-connected payments/payouts and an LLM-based support chatbot orchestrated via LangGraph, backed by a REST API over PostgreSQL.
SnapMart
A cloud-native point-of-sale system with a MERN backend, serverless image optimization via Azure Functions, and a containerized deployment on Azure Container Apps.
ShopSphere
A full-stack e-commerce platform with Stripe payments, Firebase authentication, and a full admin dashboard for product and order management.
RecipeApp
A React Native mobile app for browsing, managing, and saving recipes, with Firebase authentication and Zustand-based state management.