Md Jisan Hossen

Artifact: the official code, locked protocols, calibration freezes, and audit artifacts behind the Jev-Routed preprint — a src/ package for causal-rCM inference, controller, uncertainty, and evaluation; 50 audit, calibration, and figure-generation scripts; 34 reproducible experiment job definitions (phase 0–5c plus RQ3/RQ4); and JSON/JSONL/CSV results for every phase. The study is deliberately evidence-preserving: negative and failed experiments are reported as evidence.

DOI (v2) Project page Figures Videos

Problem: long-horizon failures of autoregressive video world models — camera drift, temporal jitter, fading spatial memory, quality decay — are described qualitatively, not measured, so runs cannot be compared.

Approach: inference-only metrics combining ORB visual odometry with Umeyama Sim(3) alignment, detrended optical-flow spectral analysis, loop-closure LPIPS, and Laplacian-variance quality trends with optional CLIP alignment. Validated against known ground-truth poses on synthetic fixtures; 24-test suite; runs on CPU.

Problem: autonomous drones must survive actuator failure, accept natural-language commands, respect social constraints, and manage energy — usually solved in isolation.

Approach: four cooperating agents around a shared 6-DOF quadrotor model — self-healing control (PPO), an LLM pilot (Mistral-7B via llama.cpp), ethical guardrails (CLIP), and a digital twin forecasting energy and wind 60 s ahead (Kalman filter). One command regenerates every figure, telemetry log, and report; 19 unit tests; CI on GitHub Actions.

Problem: it is unclear how much deep architectures buy over classical baselines when the evaluation protocol is not held constant.

Approach: a strictly identical protocol — same balanced CK+ corpus (2,450 images), same stratified split, same training loop — across KNN, SVM, a from-scratch CNN, VGG16, MobileNetV2, ResNet50, and EfficientNetB0. Results so far: ResNet50 100%, custom CNN 96.3%, EfficientNetB0 93.9%, SVM 73.1%, KNN 49.8% test accuracy.

Problem: can discrete-time filter identification be cast as learning over Z-transform pole-zero geometry, bridging classical DSP and deep learning?

Approach: pole-zero maps rendered as 64×64 images paired with 128-dim frequency-response vectors, comparing PoleZeroCNN, FreqMLP, and late-fusion FusionNet on 5-class filter classification — 99.7–100% accuracy on 750 held-out samples, with a complete 1,257-line LaTeX paper and 19 auto-generated figures.

Pairs a small local LLM (System 2) with the Jev model (System 1) to make a small model faster and more accurate — OpenAI-compatible, private, and cheap to run locally inside Kilo Code and Cline. A project page is served at j1s4nn.github.io/duomind.

A transparent proxy that routes AI coding-agent traffic to cut latency and token cost. The repository reports reducing response times from 2.4 s to 890 ms (roughly 3× faster and ~80% cheaper) through a routing layer it calls the Line J architecture, working with Kilo Code, Cline, and Claude Code.

A LangChain ReAct agent (gpt-4o-mini) that routes between FAISS-backed RAG over uploaded documents (MiniLM embeddings) and live web search, with multi-turn memory and a fully exposed reasoning trace, served through FastAPI and Streamlit.

A self-hosted service that detects duplicate questions by meaning rather than by matching words: Sentence-BERT (all-MiniLM-L6-v2) embeddings searched through a FAISS cosine index behind FastAPI, with a precision-first threshold and an automatic TF-IDF fallback.

Deep-learning analysis of DCE-MRI for breast-cancer subtype prediction, built around the TCGA-BRCA cohort with a reproducible preprocessing and evaluation pipeline.

LSTM-based forecasting on ERA5 reanalysis data, with a reproducible data pipeline and quantitative evaluation of forecast skill.

A YOLOv5 vision-aided elevator dispatch system that uses object detection to inform scheduling decisions.

An AI media toolchain integrating Real-ESRGAN super-resolution and Demucs source separation behind a unified interface.