MSc Artificial Intelligence & Robotics · Karaj, Iran
Rashin
Gholijani Farahani
I build machine learning for clinical decisions that has to survive two constraints at once: it must be cheap enough to deploy where data and hardware are scarce, and transparent enough to audit when a clinician disagrees with it. Most of my work starts from the human voice.
Open to fully funded PhD positions — Fall 2027
Research
A screening model that needs forty expensive measurements per patient is a research result, not a clinical tool. My MSc thesis treats diagnosis as a sequential decision problem: an offline reinforcement-learning policy decides, one step at a time, whether the evidence it already has is enough — or whether one more feature is worth what it costs to collect.
Because a policy that acquires features is also a policy that can be wrong, every decision is wrapped in a three-level explanation: what the model relies on globally, what drove this particular case, and a rule-based trace a clinician can read without knowing what a Q-function is.
Speech as a low-cost clinical signal
Transcript-free detection of Alzheimer's disease from spontaneous speech using handcrafted, MFCC-dominant acoustic biomarkers — no ASR stage, no transcript, no large pretrained encoder.
Cost-aware, data-efficient learning
Offline RL for active feature acquisition, so the model can learn from logged records in settings where trial-and-error on real patients is not an option.
Verification for agentic systems
A fluent model will assert unsupported claims. I am interested in pipelines where generation is cheap and verification is the expensive, non-negotiable step: claim grounding, calibrated refusal, abstention over guessing.
Publications
Transcript-Free Lightweight Detection of Alzheimer's Disease from Spontaneous Speech Using Handcrafted MFCC-Dominant Acoustic Biomarkers Under review
Interpretable, Cost-Aware Machine Learning for Speech-Based Cognitive Screening In preparation
Target: Computer Speech & Language (Elsevier) — special issue on speaker characterisation
TED: A Lightweight Explainable Framework for Emotion-Aware Arousal Estimation from Temporal Speech Energy In preparation
Target: Journal of Ambient Intelligence and Humanized Computing (Springer)
Lightweight Explainable Leaf Disease Classification Using Hybrid Features In preparation
Target: Journal of AI and Data Mining (JAIDM)
Artificial Intelligence in ADHD Diagnosis and Treatment: A PRISMA-Aligned Systematic Review IEEE
2nd International Interdisciplinary Conference on AI (IICAI 2026), Shahid Beheshti University
Adaptive Regime-Aware Portfolio Optimization Using Deep Temporal Features and HMM Regimes with Adaptive Harmony Search IEEE
15th International Conference on e-Commerce (ECDC 2026), University of Isfahan — ISC-indexed
Metaheuristic Algorithms in Video Games: Enhancing NPC Behavior in Pac-Man Using Particle Swarm Optimization
International Conference, Amirkabir University of Technology, Tehran
Code & results
Every project below is public and reproducible, and reports the number that actually matters — including when that number is bad.
MedAgent-Verify
Medical question answering where every generated claim is decomposed into atomic statements and checked against retrieved evidence. Weak retrieval triggers abstention instead of a confident guess.
Explainable leaf disease classification
87 handcrafted colour, texture and shape features into an RBF-SVM with validation-optimised fusion weights. No deep feature extractor, no GPU, no lesion segmentation — plus a calibrated rejection option.
Speaker-independent SER — an honest benchmark
The same MFCC models score near-perfectly under random splits and collapse under speaker-disjoint evaluation. Released as an argument for changing the default reporting protocol.
Attention-guided anomaly detection
A lightweight convolutional autoencoder with CBAM attention, trained on normal samples only, evaluated for both image-level detection and pixel-level localisation under a tight compute budget.
Teaching & recognition
Programming Instructor
Tehran Institute of Technology · 2023 – present
Teaching React, JavaScript and TypeScript through project-based courses and workshops. I design the course projects and assessment materials, which is where I learned that an explanation only counts if the person in front of you can rebuild the thing themselves.
Technical outreach & consulting
ELECOMP 2025 · Sharif University Career Fair · 2024 – 2025
Represented the Programming & AI Department at Iran's largest technology exhibition and advised attendees on AI and web development.
- 2026Fully funded international research fellowship in machine learning, awarded through competitive international selection
- 2025M2L Summer School — fully funded scholarship, Mediterranean Machine Learning Summer School, University of Split, Croatia
- 2025Peer reviewer, Asian Research Journal of Mathematics
- 2024Top BSc Graduate and Talent Student distinction
- 2024International Conference Recognition for AI Research, MLKD 2024
Looking for a PhD group
I am applying for fully funded PhD positions starting Fall 2027 — speech and health-oriented machine learning, interpretability, or reliable agentic systems. If that overlaps with your group's work, I would be glad to send my CV and a short research statement.