Research

Signal Lab

Atlanta, GA · generative diffusion models

• Implemented and benchmarked soft-autoregressive diffusion against standard diffusion in PyTorch, spanning 1D time-series and 2D spatiotemporal PDE (video) generation with 3D U-Net architectures on NVIDIA V100 GPUs.
• Extended a block-diffusion formulation with autoregressive rollouts for Foley sound synthesis (DCASE 2023, 4,850 clips), pushing generated audio from 4s to 14s.
• Ran ablations across ~35 configurations on the PACE HPC cluster (SLURM, multi-GPU), varying seed, width, training length and sequence length under a fairness-controlled protocol — identical total noise per position across methods, verified to ~1e-18 — establishing a monotonic dose–response and up to 75% temporal/spectral error reduction (relative L2), with multi-seed significance testing (paired t-tests, 95% CIs, effect sizes).
• Characterised the method as an inductive bias rather than a flat win: its advantage is largest under limited model capacity or training budget and converges to parity as resources grow — a claim that predicts more than a single performance number does.
• Built a six-metric diagnostic suite (Fréchet distance, autocorrelation, power spectrum, Fourier-band error, conservation drift, sample diversity) that surfaced a failure mode in FVD, which had ranked a lower-fidelity model first.
• Debugged and optimised GPU training pipelines end to end across 40+ jobs: CUDA/driver incompatibilities, mixed-precision (AMP) and TorchDynamo issues, and distributed-storage bottlenecks.

Georgia Institute of Technology, Jamali Lab

Atlanta, GA · grounded retrieval & hallucination measurement

• Built an end-to-end multi-agent RAG system (self-hosted Qwen 7B served on GPU via vLLM) that grounds scientific recommendations in a 50-document corpus and verifies every claim against its cited source, reaching 100% citation validity against a RAG baseline that fabricated references.
• Engineered a 5-tier evaluation benchmark and automated scoring pipeline, including OCR recovery of image-only PDFs, to quantify hallucination in AI-generated scientific claims.
• Extract experimental parameters from unstructured scientific literature into structured, queryable datasets for autonomous experimentation workflows.

NeurIPS 2026 workshop paper in progress, with a materials-science lab

Emory University, ViTAL Lab

Atlanta, GA · movement-disorder modelling, with Prof. Hyeokhyen Kwon

• Build an end-to-end ML pipeline to predict orthostatic hypotension in Parkinson’s patients from 3D motion-capture gait data — full-body marker trajectories from 60 patients across 3 walking tasks — spanning preprocessing, feature extraction, training and evaluation.
• Implemented and adapted DeepConvLSTM, ST-GCN and Transformer sequence models against XGBoost and Random Forest baselines, and presented the comparison to clinical collaborators.
• Engineered a leakage-safe evaluation framework: patient-grouped repeated k-fold CV so no patient appears in both splits, in-fold FDR-controlled selection over 780+ TSFRESH features, and SMOTE for class imbalance — reporting AUPRC and macro-F1 rather than accuracy, which the majority class was inflating.
• Designed a cross-medication generalisation experiment (train off-medication, test on-medication) and found that the apparent transfer could reflect subject-level confounding rather than physiological signal.

Manuscript in progress

Lumiere Education

Cambridge, MA · dimensionality reduction, mentored by a Yale PhD student

• Authored a paper benchmarking four PCA variants in Python — Standard (exact SVD), Randomised (approximate SVD), Kernel and Incremental — on UCI gene-expression RNA-Seq (800 × 16,000) and Isolet (7,800 × 620), picked so one matrix is far wider than tall and the other the reverse, measuring runtime and reconstruction error against component count over 25–100 iterations.
• The result contradicted the hypothesis: Randomised PCA bought no speed over Standard PCA and matched its reconstruction error to two decimal places, with Standard pulling ahead past ~500 components. All randomisation added was runtime variance (σ 0.57 against 0.30).
• Kernel PCA’s error turned out flat in component count — the kernel matrix preserves pairwise similarities, so extra components buy back no detail — and its runtime is set by matrix shape, ~45s tall against ~1s wide. Incremental PCA’s runtime is the flat one instead, overtaking Standard past ~400 components.