Rakib Mollah

ML Engineer at Euclido

My research spans Natural Language Processing (NLP), Computer Vision (CV), and their multimodal convergence. Specifically, I focus on advancing structured reasoning, grounding, and alignment across language models and dynamic visual representations.

Natural Language Processing & Structured Reasoning

My work in NLP centers on understanding how language models reason internally, grounding their outputs with structured knowledge, and maintaining performance under low-resource and class-imbalanced constraints.

Grounding & Reasoning Architectures

  • GRASP-ChoQ: Knowledge Graph-Augmented Stance Detection (ACL Anthology / BLP 2025)
    LLMs frequently struggle to detect subtle, indirect stances in low-resource regional contexts. We developed a Chain-of-Questions reasoning framework paired with Neo4j Knowledge Graph retrieval, improving DeepSeek R1 stance detection F1 score to 93.33 (a 40.6% relative gain).
    [paper] [pdf]

  • Enhancing LLM Reasoning via Negative Supervision and Error Reflection (Submitted to ACL ARR 2026)
    A systematic investigation into how unlearning, negative feedback signals, and self-reflection loops can guide internal model logic, allowing models to catch and correct flawed deduction trajectories prior to text generation.
    [Under Review]

Low-Resource Benchmarks & Model Efficiency

  • Moner Janala: Bengali Mental Health Counseling Dataset (Mendeley Data 2026)
    Addressed the lack of psychological NLP resources in low-resource settings by curating and structuring 625 verified counseling Q&A pairs from national archives (2009–2024) to benchmark mental health NLP models.
    [dataset]

  • SLM Optimization via Distillation & Quantization Pipelines (Submitted 2026)
    Compressed language models for Bangla affective computing (hate speech and emotion classification) using knowledge distillation and precision quantization for resource-constrained deployment.
    [Under Review]

  • Ontology-Driven Synthetic Clinical Dialogue Generation (Working Manuscript 2026)
    Formulated an ontology-constrained prompt engineering pipeline to generate structured synthetic medical dialogue in both standard and regional code-mixed Bengali.
    [Working Manuscript]

Representation Learning & Stylistic Modeling

  • Adapting Contextual Embeddings under Severe Class Imbalance (IEEE iCACCESS 2024)
    Mitigated classification failure across 568k+ Amazon e-commerce reviews by pairing RoBERTa contextual representations with class re-weighting and gradient-boosted trees.
    [paper] [pdf]

  • Deceptive Linguistic Pattern Extraction with RoBERTa (IEEE ICCIT 2023)
    Constructed a deep classifier atop RoBERTa representations to detect subtle misinformation cues on the 72k-article WELFake benchmark (99.76% accuracy).
    [paper] [pdf]

  • Autoregressive Style Replication via GPT Fine-Tuning (B.Sc. Thesis 2023)
    Fine-tuned autoregressive language models on regional literary corpora to study narrative style transfer, perplexity dynamics, and text coherence in low-resource domains.
    [pdf]

Computer Vision & Video Dynamics

This track explores model interpretability, robust real-time action recognition, and algorithmic video dynamics under unconstrained conditions.

Video Dynamics & Boundary Analysis

  • DetectVideoShotLength: Algorithmic Video Scene Cut Detection (Python Package 2024)
    Developed an open-source pipeline using color histogram divergence and Structural Similarity (SSIM) to detect scene cuts, calculate shot durations, and analyze cut distributions in unconstrained video.
    [code]

Explainable Diagnostics & Ensemble Vision

  • Grad-CAM Enhanced Pulmonary Screening from Radiographs (IEEE SMART GENCON 2023)
    Benchmarked five deep CNN architectures on 21,000+ chest X-rays, incorporating Grad-CAM saliency heatmaps to visually verify that model classifications aligned with true clinical opacities rather than spurious image artifacts (99.13% accuracy).
    [paper] [pdf]

  • Multi-Backbone Ensembles for Brain MRI Diagnostics (IEEE ICAIIHI 2023)
    Mitigated single-model variance and overfitting on structural brain MRI scans by constructing an ensemble architecture (ResNet-50, EfficientNet B7, DenseNet-121) across dual ADNI cohorts.
    [paper] [pdf]

  • Parameter-Efficient Real-Time Driver Behavior Classification (IEEE ICCIT 2022)
    Constructed a real-time vision framework combining fine-tuned MobileNet and DenseNet121 architectures to categorize 10 distinct driver distraction behaviors on the State Farm dataset (99.81% accuracy).
    [paper] [pdf]

Multimodal Convergence (Vision + Language)

The intersection of my research focuses on grounding semantic language priors within 3D/4D visual geometry and eliminating cross-modal hallucinations in Vision-Language Models.

  • Shot-GS: VLM-Guided Spatio-Temporal Reasoning for Dynamic Gaussian Splatting (Ongoing Research 2026)
    Dynamic 4D Gaussian Splatting collapses when applied to casual, multi-shot video because camera cuts break low-level optical flow and point tracking. This framework leverages Vision-Language Models (e.g., SAM 2, invariant entity tokens) to maintain cross-shot semantic identity and applies spatio-temporal reasoning across unobserved cut boundaries.
    [Ongoing Research]

  • Mitigating VLM Hallucination via DPO and Targeted Negative Sampling (Ongoing Research 2026)
    Aligning Vision-Language Models for complex document and financial reasoning by applying Direct Preference Optimization (DPO) combined with targeted negative sampling to penalize subtle visual hallucinations and numerical inaccuracies.
    [Ongoing Research]

  • Project Z: Geospatial Intelligence & Conversational Analytics (2024)
    Coupled dynamic geospatial demographic mapping with an interactive conversational assistant to synthesize regional socio-economic trends for data-driven decision support.
    [demo / repository]