Undergraduate ML researcher @ KIIT University · CGPA 9.03/10
I build things at the edge of medical image segmentation, uncertainty-aware deep learning, and agentic clinical AI. My work tends to start with a clinical problem that doesn't have a clean solution, and end with something that actually measures up against state-of-the-art. Currently in my third year, already with one published paper (Best Paper Award, ANTIC 2025) and several under review at MICCAI 2026 and Elsevier journals.
All work is first-authored unless noted.
FETALFusion · Under Review at MICCAI 2026 Dual-path CNN-Mamba encoder with resolution-aware state-space scanning for multi-domain fetal ultrasound segmentation. DSC = 0.9635, HD95 = 1.07 mm on 472 validation images. HD95 std 3.2× lower than VM-UNet.
Attention-ResUNet for Fetal Head Segmentation · Published at ANTIC 2025 (Springer LNCS) · 🏆 Best Paper, Image Processing Track Multi-scale attention gates + residual skip connections. DSC = 99.30 ± 0.14% on HC18 (n=200). Zero false-negative segmentations. 23 ms inference on RTX 3080.
BRIDGE + TCH-Net · Under Review at JNCA (Elsevier) · 📦 Model · 📊 Dataset · 📄 Preprint First formally-specified leave-one-domain-out generalisation benchmark for IoT botnet detection across 5 heterogeneous datasets (CICIDS-2017, CIC-IoT-2023, Bot-IoT, Edge-IIoTset, N-BaIoT). TCH-Net achieves F1 = 0.8296, AUC = 0.9380. Invited by Niels Rogge (Lead ML Engineer, Hugging Face) to host artifacts on the Hub; commended by the HF team for advancing reproducibility in IoT security research.
AdaptiMed · Technical Report Extended MedAgent-Pro (ICLR 2026) with guideline-grounded adaptive agentic diagnosis. +24 pp category accuracy, +26 pp broad accuracy on REFUGE2 fundus cases. −25.7% tool calls per case.
FETALFusion-v2 · In Preparation at Medical Image Analysis (Elsevier) Single-network dual-decoder CNN-Mamba for simultaneous binary segmentation + Gaussian landmark heatmaps (BPD, OFD, PS, FH). Trained on >10k images across 5 public datasets.
CAGED · Work in Progress Replacing heuristic early-exit gates in agentic diagnosis with principled Dirichlet posteriors. Guideline-RAG threshold retrieval (EGS 2020, ACC/AHA, WHO ICD-11) enables cross-domain transfer without retraining.
GPR-DE Architecture for Seismic Control of a Benchmark Cable-Stayed Bridge · Under Review at Journal of Vibration Engineering & Technologies (Springer) · 2nd Author GPR surrogate (Matérn 5/2, ARD) on 91 FEA simulations achieving LOOCV R² > 0.999; DE with UCB acquisition (κ=1.0) yielded 45% reduction in peak deck displacement under near-fault motions at ~1000× speedup over direct NLTHA-based optimisation.
ML / Deep Learning : PyTorch, scikit-learn, NumPy, Pandas, XGBoost, GPR, SciPy
Architectures : U-Net variants, Attention mechanisms, State Space Models (Mamba), BiGRU/LSTM, DS-Conv
Agentic / LLM : LLM orchestration, RAG pipelines, tool-use agentic workflows, Groq API
Research tooling : LaTeX, Weights & Biases, TensorBoard, Grad-CAM, UMAP/t-SNE, Jupyter
Languages : Python, Java, C, SQL, Bash
Infrastructure : Git, Linux, Docker
- 🏆 Best Paper Award in Image Processing Track, ANTIC 2025 (Springer LNCS)
- 🇮🇳 National Finalist at Smart India Hackathon 2025, KIIT University nominee
- 🤗 Invited Contributor at Hugging Face Hub (approached by Lead ML Engineer Niels Rogge); TCH-Net checkpoints, BRIDGE benchmark, and interactive Demo Space publicly released
Research Assistant / Internship positions in medical AI. Open to lab collaborations, especially those working at the intersection of clinical relevance and methodological rigour.
📬 ammarbhilwarawala@gmail.com 🔗 LinkedIn · Hugging Face · Kaggle · Resume
Open to a conversation about almost anything — drop a mail.