Projects
Independent projects in video, retrieval, agents, and speech, plus the earlier vision work where I started. Code for each is on GitHub.
Open-source systems I designed and built.
Multi-camera, real-time video anomaly detection. Detection and tracking (YOLOv8 + BoT-SORT) feed persistence-gated rules for falls, loitering, abandoned objects, and wrong-way motion. An LSTM trajectory autoencoder flags patterns that no rule covers.
Turns a single image into measurable 3D. Monocular metric depth is fused into 3D meshes for distance measurement, with uncertainty estimates.
Agentic retrieval for financial documents. An LLM router, grounded in the extracted table schemas, sends each question over SEC 10-K filings to hybrid sparse–dense retrieval or to sandboxed text-to-SQL, and every answer carries citations. I implemented the evaluation metrics (context precision and recall, faithfulness, relevancy) and used them to choose the retrieval configuration. In this study, a general-domain cross-encoder reranker lowered faithfulness from 0.93 to 0.77, which suggests reranking is sensitive to domain shift.
Researcher, coder, and reviewer agents built with LangGraph, with tools served over MCP. Generated code is tested in a sandbox, deterministic routing caps the number of retries, and a human approves results before anything is written.
A self-correcting retrieval-augmented generation pipeline with hallucination auditing.
A document layout and table parser built with classical computer vision.
Time-aligned video dubbing in the original speaker's voice: diarization with pyannote, Whisper transcription, NLLB translation, and XTTS-v2 voice cloning.
A voice agent that books appointments over phone calls. It streams speech-to-text, uses LLM tool calling, and speaks back with text-to-speech, handling interruptions (barge-in) and escalating to a human when needed.
The projects where I learned computer vision.
An attention-weighted ensemble of DenseNet121 and EfficientNet.
Compared FCN, attention U-Net, and DeepLabV3+.
Object detection with Faster R-CNN.
Forgery detection on the SOCOFing dataset. This work later led to my perturbation-analysis manuscript on what alteration detectors actually learn.
My undergraduate thesis at Ahsanullah University of Science and Technology: designing an arrhythmia classification algorithm using 2-D convolutional neural networks.