PhD in Computing candidate · Data Science & ML
My work asks what happens after the accuracy score: when memory is tight, arithmetic is expensive, data shifts, and results need to hold up under scrutiny.
Scroll to stress-test a model ↓
01 — THE QUESTION
I connect training-time choices to latency, memory, arithmetic cost, failure cases, and deployment constraints.
02 — RESEARCH, MADE TANGIBLE
Two small simulations turn my research interests into something you can inspect in seconds: efficient arithmetic and dependable data.
Move a learned weight. The deployment codebook snaps it to zero or a signed power of two.
Absolute rounding error: 0.120
Inspect the teaching implementation ↗Move incoming sensor data away from the training baseline and watch the monitoring decision change.
Continue monitoring. No intervention indicated.
Toy Gaussian simulation for explaining monitoring logic—not a production detector.
Inspect the monitoring project ↗03 — LIVE ML PLAYGROUND
This is a real tiny neural network running in your browser—not a video. Change the data, add noise, and watch gradient descent reshape its decision surface.
Educational synthetic data only. Results are generated locally and are not portfolio benchmark claims.
04 — INTERACTIVE SYSTEMS LAB
Move the priority slider. The recommendation changes using measured demo results from my Edge AI Benchmark Suite.
Illustrative ranking of models within the vision benchmark; measurements are environment-specific.
The compact trade-off: strong accuracy without the latency and footprint of the largest vision model.
Demo run · CIFAR-10 subset · measured values, not universal hardware claims
05 — EVIDENCE, NOT DECORATION
Explore how I structure research so another person can inspect the choices, reproduce the run, and find the limits.
Versioned configurations, fixed seeds, documented datasets, and leakage-aware splits make the comparison interpretable.
config → seed → split → runLatency, throughput, memory, model size, FLOPs, and energy proxies reveal trade-offs hidden by a single headline metric.
quality × cost × constraintsConfusion matrices, per-class metrics, drift signals, and hard examples show where the system stops being dependable.
average ≠ guaranteeCross-runtime parity checks and deployment reports keep the artifact connected to the model that was evaluated.
export → compare → approve06 — TWO-MINUTE DECISION
You own an edge vision feature. Pick the best model for each deployment brief. The technically “best” answer changes with the constraint.
Latency matters most. Some accuracy can be traded for a substantially lighter response.
07 — SELECTED SYSTEMS
Each project pairs implementation with documentation, evaluation, and explicit limitations.
Efficient AI · PyTorch
Benchmarking · Systems
Deployment · Verification
Agentic tooling · RAG
LLM reliability · Failure analysis
Decision science · Uncertainty
08 — RESEARCH RECORD
Six research outputs: four peer-reviewed publications, one preprint, and one submitted manuscript under review—spanning efficient language models, edge intelligence, medical imaging, and AI systems.
Google Scholar ↗Submitted to IEEE Transactions on Artificial Intelligence and currently under full consideration for publication.
UNDER REVIEW09 — COMPLETE PROFILE DIRECTORY
Research records, code, professional history, institutional affiliation, résumé, and direct contact.
10 — NEXT