Research, education & skills

Research experience

1. Applied ML Research, Accelx Inc.

   Jan 2025 — Present

   Machine Learning Engineer (safety-critical perception and language systems) · Dhaka, Bangladesh

   Spatiotemporal and pose-based violence recognition

   • Compared a Kinetics-pretrained R(2+1)D-18, a BERT-style Transformer over YOLO11-Pose dynamics features, and a two-stream pose–RGB cross-attention model on RWF-2000.
   • Investigated motion-aware clip mining (dense optical flow plus pose heuristics) to reduce label noise from video-level annotations.

   Grounded and abstaining multimodal and LLM systems

   • Designed a pipeline in which the LLM writes prose only from facts computed deterministically, with every citation checked by a word-overlap test whose threshold was set from observed failures (about 17% overlap for misattributed citations vs. close to 100% for correct ones).
   • Built a staged image threat-analysis cascade (OCR first, a VLM only for hard cases) with knowledge-base verification, calibrated confidence, and handoff to human review.
   • Evaluated generated text with an LLM judge, ROUGE-L, and fact recall.

   Deployed systems

   • Real-time video safety analytics (weapon, violence, fall) for live multi-camera surveillance, served at sub-100 ms latency with TorchServe.
   • AI risk & content-safety platform: fine-tuned RoBERTa classifiers, deterministic intent aggregation, and grounded, cited Llama 3.2 alerts.
   • Multimodal content & threat analysis (PaddleOCR, Qwen3-VL) and zero-shot URL categorization (BART-MNLI, Redis).

Education

1. B.Sc. in Computer Science and Engineering

   Apr 2016 — Jan 2021

   Ahsanullah University of Science and Technology · Dhaka, Bangladesh

   • Thesis: Design of an Arrhythmia Classification Algorithm Using 2-D Convolutional Neural Networks
   • Relevant coursework: Artificial Intelligence, Pattern Recognition, Digital Image Processing, Computer Graphics, Soft Computing, Data Structures, Algorithms

Technical skills