Shariar Kabir

A black-box framework that stress-tests LLMs' stated opinions using argumentative challenges, classifying responses as stable or unstable across various political topics. Applied to 12 widely-used LLMs across 19 political topics, it reveals substantial variation in stance stability—a model that is left-leaning overall can exhibit stable-right behavior on certain topics. Stability has practical implications for controlled generation and model alignment: when models are prompted or fine-tuned to adopt the opposite ideology, unstable topic stances are more susceptible to change whereas stable ones resist modification.

A critical bottleneck in mechanistic interpretability is that circuits—computational graphs revealing LLM internals—are highly information-dense and inherently prompt-specific, making large-scale behavioral analysis prohibitively laborious. While open-source tools like circuit tracer can now construct circuits for arbitrary prompts, two challenges remain: (1) raw circuits require significant human effort to cluster into interpretable supernodes, and (2) understanding an LLM's behavior on a task demands dataset-level analysis that manual inspection cannot scale to.

This project develops an LLM-based agent that takes raw circuits as input and produces actionable outputs—such as probes to detect future undesirable behaviors or surgical interventions to modify targeted behavioral tendencies. The agent-driven framework enables dataset-level circuit analysis for the first time, moving beyond single-prompt inspection toward scalable, automated interpretability. Target analyses include refusal, scheming, and internal planning.

Target venues: ICLR / NeurIPS / ICML    Deliverables: Open-source repository + technical blog post

Isolates model components responsible for specific behavioral differences through targeted ablation studies. Defines ideological depth as (i) a model's steerability and (ii) the feature richness of its internal political representations measured with sparse autoencoders (SAEs). Finds large systematic differences between Llama-3.1-8B and Gemma-2-9B, showing that refusals on benign political prompts can arise from capability deficits rather than safety guardrails—making ideological depth a measurable and tunable property of LLMs.

Curated a Bengali biomedical speech corpus for medical ASR covering two major Bengali dialects (Bengali and Sylheti). Trained and evaluated two popular ASR frameworks on a comprehensive 46-hour domain-specific corpus, achieving a WER of 8% on a symptom-focused vocabulary. This work addresses the lack of domain-specific data that limits practical healthcare ASR for underserved linguistic communities.

A system for generating domain-specific document images to fine-tune transformer models for underrepresented languages. SynthNID generates synthetic Bangla NID data that, when mixed with real data, significantly improves Key Information Extraction performance—particularly on Bengali-script fields. The system is easily extendable to generate other types of scanned documents for a wide range of document understanding tasks. (* co-first author)

A collaborative filtering-based framework for dynamic resource allocation in serverless workloads. AgnoSVD uses Singular Value Decomposition (SVD) to predict optimal resource configurations while remaining agnostic to specific function details, evaluated on AWS Lambda and Apache OpenWhisk across 99 functional workloads spanning individual functions and chains.

Existing evaluations of political bias in large language models (LLMs) typically classify outputs as left- or right-leaning. We extend this perspective by examining how ideological tendencies vary across topics and how consistently models maintain their positions, a property we refer to as stability. To capture this dimension, we propose PReSS (Political Response Stability under Stress), a black-box framework that evaluates LLMs by jointly considering model and topic context, categorizing responses into four stance types: stable-left, unstable-left, stable-right, and unstable-right. Applying PReSS to 12 widely used LLMs across 19 political topics reveals substantial variation in stance stability; for instance, a model that is left-leaning overall can exhibit stable-right behavior on certain topics. This highlights the importance of topic-aware and fine-grained evaluation of political ideologies of LLMs. Moreover, stability has practical implications for controlled generation and model alignment: interventions such as debiasing or ideology reversal should explicitly account for stance stability. Our empirical analyses reveal that when models are prompted or fine-tuned to adopt the opposite ideology, unstable topic stances are more likely to change, whereas stable ones resist modification. Thus, treating stability as a moderating factor provides a principled foundation for understanding, evaluating, and guiding interventions in politically sensitive model behavior.

In serverless computing, determining the optimal resource configurations for workloads poses significant challenges, particularly due to the cloud provider's limited visibility into workload specifics. This complexity is amplified when dealing with diverse workloads that vary in their characteristics. In this paper, we present AgnoSVD, an approach for predicting the optimum resource configuration for an incoming workload using Singular Value Decomposition (SVD). The proposed model uses collaborative filtering to extract the latent factors of the workloads and resource profiles. Therefore, the model remains agnostic to the specific details of the functions and the resource configurations. We tested our approach on well-known serverless systems like AWS lambda and Apache OpenWhisk and evaluated the system using 99 functional workloads. These workloads encompass both individual functions and chains of …

Large language models (LLMs) display recognizable political leanings, yet they vary significantly in their ability to represent a political orientation consistently. In this paper, we define ideological depth as (i) a model's ability to follow political instructions without failure (steerability), and (ii) the feature richness of its internal political representations measured with sparse autoencoders (SAEs), an unsupervised sparse dictionary learning (SDL) approach. Using Llama-3.1-8B-Instruct and Gemma-2-9B-IT as candidates, we compare prompt-based and activation-steering interventions and probe political features with publicly available SAEs. We find large, systematic differences: Gemma is more steerable in both directions and activates approximately 7.3x more distinct political features than Llama. Furthermore, causal ablations of a small targeted set of Gemma's political features to create a similar feature-poor setting induce consistent shifts in its behavior, with increased rates of refusals across topics. Together, these results indicate that refusals on benign political instructions or prompts can arise from capability deficits rather than safety guardrails. Ideological depth thus emerges as a measurable property of LLMs, and steerability serves as a window into their latent political architecture.

This study presents AmarDoctor, a multilingual voice-interactive digital health app designed to provide comprehensive patient triage and AI-driven clinical decision support for Bengali speakers, a population largely underserved in access to digital healthcare. AmarDoctor adopts a data-driven approach to strengthen primary care delivery and enable personalized health management. While platforms such as AdaHealth, WebMD, Symptomate, and K-Health have become popular in recent years, they mainly serve European demographics and languages. AmarDoctor addresses this gap with a dual-interface system for both patients and healthcare providers, supporting three major Bengali dialects. At its core, the patient module uses an adaptive questioning algorithm to assess symptoms and guide users toward the appropriate specialist. To overcome digital literacy barriers, it integrates a voice-interactive AI assistant that navigates users through the app services. Complementing this, the clinician-facing interface incorporates AI-powered decision support that enhances workflow efficiency by generating structured provisional diagnoses and treatment recommendations. These outputs inform key services such as e-prescriptions, video consultations, and medical record management. To validate clinical accuracy, the system was evaluated against a gold-standard set of 185 clinical vignettes developed by experienced physicians. Effectiveness was further assessed by comparing AmarDoctor performance with five independent physicians using the same vignette set. Results showed AmarDoctor achieved a top-1 diagnostic precision of 81.08 percent (versus physicians average of 50.27 percent) and a top specialty recommendation precision of 91.35 percent (versus physicians average of 62.6 percent).

This paper presents the development of a prototype Automatic Speech Recognition (ASR) system specifically designed for Bengali biomedical data. Recent advancements in Bengali ASR are encouraging, but a lack of domain-specific data limits the creation of practical healthcare ASR models. This project bridges this gap by developing an ASR system tailored for Bengali medical terms like symptoms, severity levels, and diseases, encompassing two major dialects: Bengali and Sylheti. We train and evaluate two popular ASR frameworks on a comprehensive 46-hour Bengali medical corpus. Our core objective is to create deployable health-domain ASR systems for digital health applications, ultimately increasing accessibility for non-technical users in the healthcare sector.

End-to-end Document Key Information Extraction models require a lot of compute and labeled data to perform well on real datasets. This is particularly challenging for low-resource languages like Bangla where domain-specific multimodal document datasets are scarcely available. In this paper, we have introduced SynthNID, a system to generate domain-specific document image data for training OCR-less end-to-end Key Information Extraction systems. We show the generated data improves the performance of the extraction model on real datasets and the system is easily extendable to generate other types of scanned documents for a wide range of document understanding tasks.

This project investigates multi-label medical code classification by training linear probes on Large Language Model (LLM) activations. We extract layer-wise attention head activations from medical-domain LLMs and use Ridge regression classifiers to predict relevant medical disciplines from clinical descriptions. The approach enables interpretable analysis of which model components are most informative for medical domain classification tasks.

Recognizing the inherent bias of most LLMs towards European languages and ethnicities and the low resources of structured Bengali data, I focused on refining open-source models like LLaMA using different parameter-efficient fine-tuning (PEFT) techniques (e.g., Adapter injections and LoRA). Successfully fine-tuned LLaMA for Bengali mental health consultation using QLoRA, resulting in a more optimized model that can be served on low GPU memory—ensuring equitable access to healthcare technologies across diverse linguistic communities.

ASR system for understanding medical symptoms spoken by patients in Bengali. Trained DeepSpeech from scratch on audio collected via a consented data collection portal, then finetuned for noisy environments using 13 domain augmentations. Switched to a Whisper (tiny) model finetuned on the BanglaASR corpus (Bangla Mozilla Common Voice), achieving a WER of only 8%—enabled by the limited vocabulary of symptom terms.

A system leveraging Vision-Language Models (VLMs) to assist users in performing exercises correctly by comparing their execution against reference videos of expert demonstrations. Uses frame-level visual and motion comparison integrated with language-based feedback to generate natural language guidance that helps users improve their form and reduce injury risk.

A disease recommendation system using ensemble classifiers trained on synthetic patient data reflecting real-world demographics. The data generator accounts for risk factors, family history, and medical history. The classifier uses a multi-layer pipeline: first predicting disease probabilities from symptoms, then filtering using an ethnicity-based prevalence look-up table, and finally making a final prediction using the patient's risk factors.

Species tree estimation from gene trees is crucial in phylogenetics. Quartet-based techniques like ASTRAL, QMC, and wQFM are widely used, but struggle with scalability on large datasets. This project redesigns wQFM for distributed execution using Apache Spark, analyzing the scalability and performance gains on large-scale phylogenomic inputs.

A distributed music playlist system implemented using the three-phase commit (3PC) protocol to guarantee consistency across multiple nodes. The playlist is an unordered set of song name and URL pairs, maintained with full ACID compliance across two or more devices.

Fine-tuned the pretrained Donut document transformer model for Key Information Extraction (KIE) on Bangla National ID cards using data generated by SynthNID. Used a mix of real and synthetic data, with the addition of synthetic data yielding significant performance improvements—especially on Bengali-script fields.

A collection of applied object detection tasks built on YOLO, each fine-tuned for a specific domain:

• Licence plate detection — fine-tuned YOLOv5 to detect plates across vehicle classes in Bangladeshi CCTV footage. Colab
• Card region extraction — locating and cropping identity card regions from raw photographs for downstream processing.
• Engineering drawing checker — detecting standard symbols (e.g. north sign) in scanned engineering drawings to automate compliance checks.

Extraction of named entities (NE) from banking chatbot messages—beneficiary names, transfer amounts, account types, account numbers, etc. Instead of a single model, we used a recipe of approaches: BERT for beneficiary name extraction, RegEx for transfer amounts and account numbers, and lookup tables for account type extraction. This minimized the training burden for Bengali, a low-resource language.

Bangladesh's pioneering voice-based AI chatbot for seamless banking activities, serving hundreds of thousands of real users at Agrani Bank—one of Bangladesh's largest state-owned banks. Powered by Bengali ASR and a finetuned NLU engine for natural language-driven fund transfers and inquiries, making banking services accessible to customers with limited digital literacy.

Analyzes real-time facial movements and blinking, and requires the user to perform specific facial actions during the eKYC authentication process to ensure the presence of a live person. Developed for deployment on mobile devices such as smartphones.

Audio data collection portal for large user bases. Built with a React frontend and Python-Flask backend; metadata stored in PostgreSQL, object storage in S3, and full authentication via AWS Cognito. Supports priority-based or demographic-filtered collection, useful for gathering medical recordings segmented by symptom, age, or gender.

A portal for showcasing and demo-testing AI services. Authentication and authorization built using Keycloak and Google identity provider. New clients can sign up with their email and receive a limited credit allocation for trying the services.

A multiplayer microcontroller game featuring two player-controlled bars on either side of two connected dot matrices. Powered by a single Atmega32 microprocessor, with controllers built using MPU-6050 accelerometer & gyro sensors.

A ray tracer that renders spheres, planes, and triangles with textures and shadows by tracing light paths through an image plane. Implements the Phong Lighting Model and recursive reflection for photorealistic lighting effects.