animesh = {
"role" : "AI Researcher & MLOps Engineer",
"education" : ["MSc Advanced Computer Science @ Newcastle University (2025–26, on track for Distinction)",
"B.Tech CSE (AI Specialization) @ AKTU, India"],
"research" : ["Medical Image Analysis", "Computer Vision",
"Generative AI", "Clinical AI Safety",
"Uncertainty Quantification"],
"goal" : "Funded PhD in AI/Medical Imaging (Germany, 2026) — targeting TU Munich, DKFZ, FAU",
"location" : "Newcastle upon Tyne, UK 🇬🇧",
"visa" : "Eligible: UK Graduate Visa & Netherlands Orientation Year Visa (2026)",
"fun_fact" : "I turn retinal scans into decisions — one attention map at a time 🔬"
}
PhD-level independent research — multi-class retinal fluid segmentation with clinical uncertainty triage
- Architecture: Dual AttentionTransUNetL ensemble (EfficientNetV2L encoder, 127M params each) + Transformer bottleneck (d_model=512, 16 heads) + 4× Attention Gates + Source-Adaptive BatchNorm
- Dataset: 4 independent OCT sources — DUKE DME, AROI, UMN AMD, UMN DME (4983 training slices)
- Results: V2L val Dice 0.784 ± 0.006 across 3 seeds · IRF 0.916 · SRF 0.856 · PED 0.581
- Novel: UCUS — Uncertainty-Weighted Clinical Urgency Score (Monitor / Review / Urgent triage)
- Clinical Safety: Uncertainty 1.34× higher at inter-grader disagreement pixels (p=3.77e-05) · SRF volume r=0.778 · PED volume r=0.841
- Deployment: INT8 quantised (510MB → 132MB, 3.9×) · ONNX export · Streamlit dashboard · FastAPI endpoint
- Targeting: arXiv preprint + OMIA 2027 Workshop at MICCAI
PyTorch EfficientNetV2L TransUNet MC Dropout ONNX Streamlit FastAPI HuggingFace
Production-grade clinical AI system for automated retinal disease detection
- Architecture: EfficientNetV2L + 4× Multi-Head Attention + Learnable Positional Encoding + XGBoost hybrid head
- Dataset: Kermany et al. (84K OCT images — CNV, DME, DRUSEN, NORMAL)
- Results: 5-seed validated · 95.43% ± 0.27% Accuracy · Macro AUC 0.9941 ± 0.0006 · ECE 0.0024
- RETFound comparison: Matches 303M-parameter foundation model while being 5× more stable across seeds
- Clinical Safety: Mahalanobis OOD Detection · MC Dropout · Temperature Scaling · Grad-CAM · SHAP
- Deployment: ONNX 237MB · ~62.9ms CPU · Streamlit + Gradio + FastAPI on HuggingFace
TensorFlow EfficientNetV2L XGBoost Optuna SHAP Streamlit HuggingFace
End-to-end clinical pipeline connecting both retinal AI projects
- Stage 1: Classification (CNV / DME / DRUSEN / NORMAL) via ONNX inference
- Stage 2: Fluid segmentation (IRF / SRF / PED) with live ONNX dual ensemble
- UCUS clinical triage score computed end-to-end from raw scan to urgency band
Research-grade plant pathology classification — 38 diseases, 54,306 images
- Results: 99.57% Test Accuracy · 99.48% Macro F1 · McNemar p = 3.27 × 10⁻¹⁸²
- Architecture: EfficientNetV2S · two-stage transfer learning · TFLite float16 (~45 MB)
TensorFlow EfficientNetV2S TFLite Streamlit UMAP HuggingFace
Automated content creation pipeline with RAG + fine-tuned Llama-2
- ~70% reduction in manual writing time · LangChain RAG · vector store retrieval
LangChain RAG Llama-2 Python
| Certification | Issuer | Badge |
|---|---|---|
| Oracle Generative AI Professional | Oracle |  |
| AWS Solutions Architect | Amazon Web Services |  |
| Azure Fundamentals (AZ-900) | Microsoft |  |
| Oracle AI Vector Search | Oracle |  |
🔗 View all credentials on Credly →
🔷 IBM — AI/ML Intern (Summer 2025)
LLMs & Transformer architectures — fine-tuning, deployment, and prompt engineering at enterprise scale
🔷 IIT Kanpur — Deep Learning Intern (May–June 2023)
Retinal disease detection with DL + AWS cloud infrastructure
🔷 MedTourEasy — Data Analyst (October 2022)
Healthcare data analytics and reporting
- 🔬 OCT Fluid Segmentation — arXiv preprint submission (targeting cs.CV / eess.IV)
- 🔬 OCT Classification — arXiv preprint submission
- 🎓 PhD Applications — targeting funded positions at TU Munich · DKFZ · FAU (Sept/Oct 2026)
- 📖 MSc Dissertation — Newcastle University (2025–26)
Medical Image Analysis Computer Vision Generative AI Clinical AI Safety
Uncertainty Quantification Explainable AI (XAI) Transformer Architectures MLOps
