Miray Sude Mutlu

• Built v4, a dual-sequence binding model: shared-weight ESM-2 (35M) encoders for TCR CDR3β and epitope, Atchley physicochemical features, cross-attention, an HLA pseudo-sequence encoder and a V/J gene channel; the gene channel alone adds +0.1335 macro-AUROC in ablation.
• Engineered an epitope-strict data pipeline over 8 databases (VDJdb, McPAS-TCR, NeoTCR, CEDAR, MIRA, ImmRep23 and others): 314,421 training rows across 1,053 unique epitopes, with leakage-free CDR3β-strict splits and HLA-aware negative sampling.
• Ranked first of six open models on a frozen 15,715-row held-out set (macro-AUROC 0.686); on the ImmRep23 out-of-distribution set it stays near random (0.502), which we report openly as the field-wide generalisation ceiling.
• Wrapped predictions in a three-engine decision-support platform (Next.js, FastAPI, Docker): temperature calibration with Mondrian conformal abstention, TCGA-GBM evidence fusion, a 20,431-protein off-target safety scan and a Turkey-specific HLA panel reaching 73.15% coverage with five epitopes.
• Stress-tested the platform end to end with 3,200 real CDR3β sequences from 8 GBM patients, with zero crashes.

• Python
• PyTorch
• ESM-2
• HuggingFace Transformers
• Scikit-learn
• FastAPI
• Next.js
• Docker
• WandB
• RunPod (H100 GPU)

• Fused 6 pharmacological databases (DrugBank, SIDER, FAERS, DailyMed, ChEMBL, DrugBank Pathways) into a unified dataset of 1,409 drugs × 975 features via a multi-source ETL pipeline with NLP-based SOC extraction from FDA SPL documents.
• Built a two-stage ML architecture: FAERS-tier specialist ensemble (XGBoost / Balanced RF / GNN-GAT hybrid) for binary ADR risk scoring (0–100), plus Binary Relevance RF for 17 WHO MedDRA SOC classes (F1 Macro 0.782, best class Skin F1 0.955).
• Implemented ADR Burden Score, patient personalisation (Charlson CCI + Beers Criteria), guideline-anchored alert engine (ADA 2024, KDIGO 2021, ESC/ESH 2023), and ATC + SOC-filtered alternative drug recommendations with ~22,000 Turkish brand-name lookup.
• Deployed React 18 / TypeScript clinical interface with SHAP feature attribution, per-SOC probability charts, and a parallel multi-agent LLM assistant: 4 models (Llama 3.3 70B, Llama 3.1 8B, GPT-4o Mini, Gemini 2.0 Flash) queried concurrently via Promise.all with consensus synthesis by a GPT-4o Mini judge.

• Python
• Scikit-learn
• XGBoost
• PyTorch
• RDKit
• React
• TypeScript
• FastAPI
• PostgreSQL
• SHAP
• Optuna
• Groq API
• OpenRouter

• Conducted feature engineering on 871 MRI samples with 10,734+ radiomic features; evaluated 8 feature selection methods (LASSO, ANOVA, Chi², Mutual Information, Wilcoxon, MRMR, RF Importance, hybrid).
• Benchmarked 64 model configurations (8 feature selection × 8 classifiers) using 5-fold stratified cross-validation with SMOTE for class imbalance (523 control vs. 348 Alzheimer's).
• Achieved ROC-AUC 0.855 (LASSO + Logistic Regression) with bootstrap-validated 95% CI [0.839–0.871]; generated automated performance visualisations.

• Python
• Scikit-learn
• Pandas
• SMOTE
• LightGBM
• XGBoost

• Consolidated raw clinical obstetric patient data and programmatically resolved Excel locale-induced date corruption, out-of-biological-range laboratory anomalies, and duplicate records.
• Engineered 48 new features from repeated laboratory measurements using clinical domain knowledge; managed 14.6% class imbalance in the multi-rule OR-defined target variable using SMOTE.
• Built a fully leakage-safe end-to-end pipeline; trained 9 models including XGBoost, LightGBM, CatBoost, and Random Forest via 5-Fold Cross-Validation, achieving AUC 0.890.
• Optimised the decision threshold using the Youden Index, reducing the number of high-risk cases missed by the model by 18%.
• Conducted SHAP analysis to make model decisions interpretable for clinicians (explainable AI); integrated findings into a manuscript draft targeting peer-reviewed publication in 2026.

• Python
• Pandas
• NumPy
• Scikit-learn
• XGBoost
• LightGBM
• CatBoost
• Matplotlib
• SHAP
• SMOTE

• Designed and implemented a production-level backend architecture: user management, role-based access control, and JWT-based authentication.
• Optimised PostgreSQL schema and inter-service API contracts for scalability and maintainable integration.
• Established clear API documentation to stabilise the development and testing pipeline across the team.

• Java
• Spring Boot
• PostgreSQL
• RESTful API
• JWT

• Built an end-to-end pipeline that converts a single text prompt into a multi-speaker audio podcast: GPT-2 generates a structured two-speaker dialogue (context window: last 4 lines, temperature 0.75), then Suno Bark (small) synthesises each line to 22,050 Hz WAV.
• Implemented automatic language detection (EN/FR), per-speaker voice differentiation, and Pydub-based audio concatenation with 300ms silence padding, exported as a single final_podcast.wav.
• Exposed the full pipeline through a Gradio web interface with real-time progress feedback, built-in audio preview, and one-click download.

• Python
• HuggingFace Transformers
• GPT-2
• Bark TTS
• Gradio
• Pydub
• NumPy

• Designed and benchmarked ML and deep learning models for real-time network attack classification on the UNSW-NB15 benchmark dataset.
• Applied cross-validation and ROC-based evaluation to identify the most balanced model configuration across attack categories.
• Compiled findings into a paper presented as first author at IEEE ASYU 2025, published in IEEE Conference Proceedings. DOI: 10.1109/ASYU67174.2025.11208415

• Python
• Scikit-learn
• Keras
• Pandas
• Matplotlib

• Built a full-stack mobile travel journal (React Native + Expo) with Firebase Authentication and Firestore for real-time cloud storage; supports complete CRUD operations for travel memories with multi-photo uploads, GPS tagging, and personal journaling.
• Implemented an interactive map view (react-native-maps) with custom markers, a digital passport that auto-generates country flags from ISO 3166-1 codes via Nominatim reverse geocoding, and a statistics dashboard tracking countries and cities visited.
• Solved real-time state synchronisation between screens using React Navigation's useIsFocused hook, and built a custom country-name-to-ISO mapping layer for consistent flag emoji rendering across locales.

• React Native
• Expo
• Firebase Auth
• Firebase Firestore
• react-native-maps
• Nominatim API
• expo-location
• expo-image-picker

A data-driven smart transportation system for Istanbul's 5.5 million daily passengers — tackling fuel waste, lack of dynamic routing, and accessibility gaps through ML, graph theory, and an inclusive NLP assistant.

• Trained a Gradient Boosting model on 38M+ rows of IETT trip data (Accuracy 0.8997, F1 0.8607) to predict low-occupancy trips; recommended replacing articulated buses (160 cap.) with standard buses (94 cap.) during off-peak hours — estimated annual saving of ₺199,131,896.
• Built a dynamic route management module that detects real-time failures and cancelled trips, automatically suggests alternative routes, and redirects nearby available vehicles to critical points.
• Developed an inclusive NLP + Groq LLM assistant (voice + chat) with intent parsing, slot filling, and GTFS API integration — enabling GPS-based nearest stop detection and real-time trip status queries for elderly and disabled users.

• Python
• Gradient Boosting
• Streamlit
• Flask
• Groq LLM
• NLP
• GTFS API
• Pandas

A clinical decision support system (CDSS) triggered the moment a physician prescribes a drug — scanning the patient's full lab history, chronic conditions, and active medications to generate a personalised risk score.

• Fused a hybrid decision engine (Rule×0.50 + LightGBM×0.30 + NLP×0.20) across 26 clinical rules for 12 drug groups (NSAIDs, Metformin, ACE inhibitors, Statins, Anticoagulants…), benchmarked against 4 models — LightGBM selected with AUC 0.9664, F1 0.8964 on 39,000 patients and 45 features.
• Queried 71M+ lab records via DuckDB with window-function-based GFR trending, HbA1c unit normalisation, and polypharmacy detection; validated drug-lab interactions retrospectively using Wilcoxon Signed-Rank tests (Diclofenac → creatinine ↑, Statin → CK ↑, ACE → K⁺ ↑, Loop diuretic → K⁺ ↓, all p < 0.05).
• Built a Streamlit dashboard with colour-coded risk cards (RED/YELLOW/GREEN), NLP-generated pharmacological mechanism explanations (evidence level A/B/C via Nexpath API), voice input/output (Web Speech API, tr-TR), and ATC-filtered safer alternative drug suggestions. Achieved 7/7 correct classifications across 6 test patients.

• Python
• LightGBM
• XGBoost
• Streamlit
• DuckDB
• SHAP
• SMOTE
• Nexpath API
• Web Speech API

A data-driven smart city mobility ecosystem built in 32 hours using real Istanbul open data. Focuses on comfort and sustainability rather than just speed — making public transit attractive by predicting crowding before it happens.

• Trained crowding prediction models (CatBoost, LSTM, XGBoost, Random Forest, SVM, DNN) on hourly passenger data from the IBB Open Data Portal, achieving 99.83% accuracy for station occupancy forecasting.
• Built route optimisation using Ant Colony Optimisation (ACO) and graph theory with real road distances via OSRM (Open Source Routing Machine), enabling comfort-score-based routing that avoids peak crowding.
• Developed a dual-product system: a React Native mobile app for passengers (comfort-based routing, Istanbul transit hacks) and a React.js + Leaflet.js web dashboard for IBB operators (real-time station monitoring, automatic extra-service interval recommendations via FastAPI).

• Python
• CatBoost
• LSTM
• XGBoost
• FastAPI
• React.js
• React Native
• Expo
• Tailwind CSS
• Leaflet.js
• OSRM
• TensorFlow
• Scikit-learn