// field notes
Research
I’m in three labs right now, working on things that don’t look related from the
outside: generative audio, a retrieval system for scientific papers, and gait data
from patients. What I keep coming back to is evaluation. Working out whether a model
is actually better, or just looks better, is usually the harder half.
Machine Learning Research Intern
Jun 2026 – Aug 2026
Signal Lab
Atlanta, GA · diffusion models for generative audio
- Implemented soft-autoregressive and standard diffusion models in PyTorch/NumPy for Foley sound synthesis (DCASE 2023, 4,850 clips), extending a block-diffusion formulation with autoregressive rollouts from 4s to 14s outputs.
- Designed a fairness-controlled comparison enforcing identical total noise per datapoint across methods (verified to ~1e-18), then ran multi-seed significance testing (paired t-tests, 95% CIs, effect sizes) isolating a 43% temporal/spectral error reduction (p<0.01) in structured regimes.
- Built a six-metric evaluation suite (Fréchet distance, autocorrelation, power spectrum, diversity, nearest-neighbour realism, distributional divergence) that surfaced a failure mode in the standard FAD/FVD metric, which had ranked a lower-fidelity model first.
- Ran 40+ multi-GPU SLURM jobs on the PACE HPC cluster, resolving CUDA/driver and storage-quota issues across parallel training and evaluation runs.
Machine Learning Research Assistant
May 2026 – Present
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 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 measure 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
Machine Learning Researcher
Aug 2025 – Present
Emory University, ViTAL Lab
Atlanta, GA · movement-disorder modelling, with Prof. Hyeok Kwon
- Build production-style ML/DL pipelines for movement-disorder modelling with Prof. Hyeok Kwon.
- Developed orthostatic hypotension detection models on full-body gait motion-capture data from 60 patients across 3 walking tasks, using feature-based and deep time-series approaches on Emory BMI servers.
- Evaluated XGBoost and Random Forest against DeepConvLSTM, ST-GCN, and Transformer baselines, reaching 0.70–0.80 accuracy across clinical settings.
Ongoing work toward authorship
Methods & tools
Research methods
Experimental designSignificance testingAblation & baseline studies
Evaluation benchmark designReproducibilityFeature engineering
Modelling
PyTorchDiffusion & generative modelsTransformers
Time-series modellingGraph neural networksReinforcement learning
Computer visionscikit-learnXGBoost
Generative AI
LLMsRetrieval-augmented generationMulti-agent systems
vLLMEmbeddingsHybrid search
Hallucination & grounding evaluationHugging Face
Infrastructure
SLURMHPC & multi-GPU trainingCUDA
GPU model servingLinuxDockerGit
AWSGCP