Sajad Abavisani
Ph.D., Georgia Tech · Causal AI Researcher
Building the causal layer of AI — algorithms that are not just accurate, but trustworthy enough to deploy in neuroscience, finance, and healthcare. ICLR 2023 Oral · NeurIPS 2026 · 10,000× speedup in causal discovery · applied to fMRI from 311 people with and without schizophrenia. Lately: whether transformers can learn causal structure directly from slowly sampled recordings, and agents on local LLMs that research markets every morning.
Actively seeking Research Scientist, Applied Scientist, and Quantitative Researcher roles. U.S. permanent resident (green card holder) — no sponsorship required.
Education
View Full Resume →Ph.D. in Electrical and Computer Engineering
Aug 2020 – Aug 2026Georgia Institute of Technology
GPA: 3.8/4.0
Select coursework: Mathematical Foundations of ML, ML with Limited Supervision, Conversational AI, Convex Optimization, Statistical ML, Information Theory, Random Processes
M.S. in Electrical and Computer Engineering
Aug 2020 – Aug 2023Georgia Institute of Technology
GPA: 3.9/4.0
B.S. in Electrical Engineering
Sep 2014 – Sep 2019Ferdowsi University of Mashhad
GPA: 4.0/4.0
Research & Publications
Google Scholar →- RnR: A Meta-Solver for Causal Discovery in Undersampled Time Series – NeurIPS 2026, accepted (45% higher edge F1 than four widely used methods)
- Recovering Directional Brain Networks under Temporal Undersampling: An Application to Schizophrenia – Under review at Network Neuroscience (bioRxiv preprint, 2026)
- GRACE-C: Generalized Rate Agnostic Causal Estimation via Constraints – ICLR 2023 (notable top 25%, oral presentation)
- Causal Learning through Deliberate Undersampling – Proceedings of the Second Conference on Causal Learning and Reasoning (2023)
- Reducing Causal Illusions through Deliberate Undersampling – NeurIPS 2022 Workshop (2022)
- Greedy AutoAugment – Pattern Recognition Letters Journal (Elsevier, 2020)
- Radiologist-Level COVID-19 Detection Using CT scans – (2020)
- Causal Graph Recovery in Neuroimaging through Answer Set Programming – Preprint
- ION-C: Integration of Overlapping Networks via Constraints – Under review at ICLR 2027 (arXiv preprint, 2024)
Featured Research
GRACE-C: Generalized Rate Agnostic Causal Estimation
A breakthrough method for causal discovery in undersampled time series data, cutting a benchmark from 17 hours to 6 seconds. ICLR 2023 (notable top 25%, oral presentation).
Learn More →
RnR: Meta-Solver for Causal Discovery
An ASP-based meta-solver that refines outputs from any causal discovery algorithm (PCMCI, FASK, MVGC, GIMME), raising edge F1 by 45% over those methods. Accepted at NeurIPS 2026; its schizophrenia fMRI application is under review at Network Neuroscience.
Learn More →
Brain-as-Language: Transformers for Causal Discovery
Can a learned model take over from an exact solver? A discrete brain vocabulary that is provably closed under slower sampling, and a time × region transformer trained on ~1M simulated recordings that scores directed links at 0.71–0.77 AUROC. Paper in preparation.
Learn More →
Causal Learning through Deliberate Undersampling
Groundbreaking work showing that measuring data less frequently can actually provide more information about causality by reducing ambiguity in causal relationships.
Learn More →
Solution-Set Transformer for Brain Classification
A Graph Transformer + Set Transformer architecture that learns diagnostic biomarkers from multi-graph RASL output, classifying healthy vs. schizophrenia subjects directly from fMRI connectivity.
Learn More →
Radiologist-Level COVID-19 Detection
Novel DECAPS architecture using detail-oriented capsule networks for automated COVID-19 diagnosis from CT scans, achieving radiologist-level accuracy.
Learn More →Technical Projects
View All Projects →
Morning Brief: Market Research on a Local LLM Council
An agent that researches a live portfolio every trading day on a home server: keyless quotes and news, a two-model local council (Gemma 4 triage, Qwen3.5-122B verdict) that keeps only unanimous signals, earnings dates to Google Calendar, and an 8 AM email — seven checkpointed, resumable stages scheduled by OpenClaw, with a GPT-5.5 chairman council service for open-ended questions.
View Details →
Local LLM Monitor
A native macOS observability dashboard for on-device models: loaded Ollama models and their unified-memory footprint, live CPU and memory pressure, a metadata-only proxy that records token counts and timing without ever reading prompts, and an agent/task tree published by a small CLI.
View Details →
LLM-Guided Causal Graph Selection
LoRA/QLoRA fine-tuning of Mistral-7B on synthetic gunfolds data to select ground-truth DAGs from RASL equivalence classes — combining chain-of-thought reasoning with learned domain priors.
View Details →
Quantfolio: Portfolio Signal Engine
A three-tier attention system (morning brief → hourly digest → push alert) with a 4-axis scoring pipeline: materiality, portfolio-weight impact, SHA-256 novelty dedup, and urgency. Claude API synthesis, SQLite persistence, ntfy.sh phone push, and a curated sector theme graph — all deployable via cron/launchd.
View Details →
Greedy AutoAugment
Efficient image data augmentation using greedy search strategy, achieving comparable accuracy on CIFAR-10, CIFAR-100, and SVHN while using 360x fewer computational resources.
View Details →
ION-C: Integration of Overlapping Networks
Computationally efficient algorithm that builds unified causal models from separate, partially overlapping datasets, scaling to graphs with up to 25 nodes with 99.55% accuracy. Under review at ICLR 2027.
View Details →
RnR: Causal Discovery Meta-Solver
Answer Set Programming meta-solver that recovers true causal structure from undersampled, noisy neuroimaging time series — applied to fMRI from 311 people with and without schizophrenia. Accepted at NeurIPS 2026.
View Details →Skills & Expertise
View Full Resume & Skills →Languages
Python (primary), C++, Java, R, SQL, Bash
ML & Statistics
PyTorch, TensorFlow, scikit-learn, NumPy, SciPy, pandas, statsmodels; hypothesis testing, bootstrap, false discovery rate control
Causal Inference & Time Series
Causal discovery, Granger causality, VAR, PCMCI, multivariate time-series modeling, evaluation design with pre-specified baselines and negative controls
LLM Systems
Hugging Face Transformers and PEFT (LoRA/QLoRA), retrieval-augmented generation, multi-agent pipelines, evaluation harnesses, local inference (Ollama), FastAPI
Graphs & Optimization
Graph algorithms, combinatorial optimization (answer set programming, clingo)
Infrastructure
Linux, SLURM HPC clusters (A100/H100 GPUs), Docker, Git, GitHub Actions CI/CD, pytest, AI-assisted development (Claude Code)