Syed Affan Daimi

MS in Computer Science, IIIT Hyderabad — Language Model Evaluation & Reliability

I’m a Master’s student in Computer Science at IIIT Hyderabad (Transcript here), where I have studied how LLM-as-judge systems behave across different axes of news summarization, and how architectural choices (activation functions, in SG-Blend) affect the robustness of learned representations. Alongside this I’ve built a 148M-parameter trilingual LM from scratch (TTGemma) and applied LLMs to structured extraction problems, from teacher-training question generation (VidhyaDojo) to parsing SEC 10-K filings.

Before this I worked across computer vision and applied data science: a ViT-based object detector for autonomous driving, demand forecasting at Razor Group in Berlin, and CLIP-based image search at fellowship.ai, TA data visualization at IIIT-H, which is also where my thesis work in visual analytics for dynamic networks (DyNetSpin, EuroVis 2026) comes from.

Publications

Glass Doors: Competency as a Behavioral Dimension for Interpreting Agent Failure.Under Review
Syed Affan Daimi, Gaurav Sarkar, Kamalakar Karlapalem · Submitted to NeurIPS 2026 Workshop IAB
Mask-Pattern Position Augmentation: Training-Free Long-Context Extrapolation for Diffusion LLMs.Under Review
Gaurav Sarkar, Syed Affan Daimi · Submitted to NeurIPS 2026 Workshop DiffuLM
SG-Blend: Learning an Interpolation Between Improved Swish and GELU for Robust Neural Representations.Preprint
Gaurav Sarkar, Syed Affan Daimi, Jay Gala, Subarna Tripathi · arXiv:2505.23942
DyNetSpin: Analysis of Dynamic Networks via Spiral Layouts.Posters & Demos
Syed Daimi*, Garima Jindal, Kamal Karlapalem · EuroVis 2026
A Scalable Data-Driven Framework for Systematic Analysis of SEC 10-K Filings Using Large Language Models.Preprint
Daimi, S. A.* · arXiv:2409.17581
Leveraging Machine Learning Models For Proactive Disaster Forecasting.
Daimi, S. A., Askari, A. A., et al. · Proceedings of ICAET 2025
Performance Efficient and Fault Tolerant Approximate Adder.
Iqbal, A., Daimi, S. A., & Chari, K. M. · Journal of Electronic Testing, 39(5), 571-582

Select Writing

Enhance Your LLM Agents with BM25: Lightweight Retrieval That Works ↗

How lightweight BM25 retrieval can outperform heavier setups when grounding LLM agents.

LLMsRetrievalAgents

Creating Custom Dashboards with Vizro: A Comprehensive Guide ↗

A walkthrough of building customizable, filter-driven dashboards using McKinsey's Vizro library.

Data VisualizationPythonVizro

Deploy Your Streamlit App to Heroku in 3 Easy Steps ↗

A short, practical guide to shipping a Streamlit app to production on Heroku.

StreamlitDeployment

Linear Classifiers on Steroids: The Tiny Recursive Models ↗

Code-first Breakdown of the Probabilistic TRM Model, from step 1.

PyTorchDeep Learning ArchitecturesLLMs

Select Projects

10-K Filings Rating System ↗

LLM-driven pipeline that extracts SEC 10-K filings and rates companies across confidence, environment, innovation, and people dimensions.

PythonLLMsStreamlitCohere

DyNetSpin ↗

Visual analytics tool for interpreting community evolution in large, time-varying networks using a novel spiral-layout algorithm.

PythonD3.jsVisualization

HPC Workflow Toolkit ↗

CLI suite automating SLURM node allocation, SSH tunneling, and remote VS Code Server setup on a university HPC cluster into one command.

ShellSLURMDevOps

Contact

Best reached by email — saffand03@gmail.com. Links to GitHub, LinkedIn, and other profiles are above.