I like building systems that are easier to trust because they are easier to debug.

Most of my work sits somewhere between applied ML and systems engineering: edge demos on QEMU, researcher-facing ML tools, repeatable pipelines, and test harnesses that catch problems before people do. I care about logs, profiling, instrumentation, weird failure modes, and making models behave like real software.

• Currently at Rutgers University pursuing an M.S. in Computer Science.
• Previously worked on bird-species classification at scale and fatigue prediction from wearable data.
• Right now I am pushing harder into edge AI, verification-minded tooling, and reliable ML workflows.
• Outside core engineering work, I am also an Adobe Student Ambassador, running workshops and creating technical content on campus.

Zephyr + QEMU + TensorFlow Lite Micro + CMSIS-NN

This is the project that feels most like the direction I want to keep pushing.

• Built an ARM Cortex-M edge AI demo with RTOS threading, UART JSON output, runtime diagnostics, latency profiling, and stack monitoring.
• Used QEMU to make embedded debugging less painful and more repeatable.
• Tracked down and fixed an inference-thread stack overflow by actually measuring stack behavior instead of guessing.

GitHub Actions + Python + QEMU

I do not like green checkmarks that tell you nothing. This work was about making failures readable.

• Added smoke-test validation so firmware builds had to produce expected structured output.
• Built log-driven assertions and deterministic checks instead of relying on "it compiled, so probably fine."
• Kept the workflow usable both in CI and in local development.

Angular + .NET + Qdrant + Hugging Face

This one is less embedded, but it shows the same pattern: build the system, then make it inspectable.

• Turned document retrieval into a full-stack workflow with semantic search, evaluation, and visibility into quality and latency.
• Improved retrieval efficiency by 70% on internal benchmarking.
• Focused on chunking, observability, and traceable changes rather than hiding the pipeline behind a single demo screen.

September 2025 - Present

• Create technical and creative content around Adobe Express and run workshops for the Rutgers student community.
• Useful side effect: it keeps me good at translating technical work for real people.

January 2025 - May 2025

• Built a classification workflow for a 710GB bird dataset covering 1,150 species, with dataset ingest, validation, and repeatable experiment runs.
• Delivered a researcher-facing CLI/GUI flow so the model could actually be used, not just trained.
• Spent a lot of time on noisy labels, confusing classes, and the practical side of making results inspectable.

October 2024 - January 2025

• Iterated a fatigue prediction pipeline through 20 versions, tightening preprocessing and evaluation on noisy wearable data.
• Reduced MAE by 78% from 1.17 to 0.25 using a PyTorch ordinal model with CORAL-style learning.
• Treated the work like engineering, not leaderboard chasing: compare changes, track regressions, and document what actually moved the metric.

• Shipping demos that are usable by someone other than the person who built them.
• Instrumentation before guesswork.
• Clean experiment flow over chaotic trial-and-error.
• Projects where software reliability matters as much as model accuracy.
• Tooling that makes debugging faster the second time, not just the first time.

• Patel, V., Shah, K., & Joshi, K. (2025). OpenPose, PoseNet and MoveNet: The evolution of deep learning methods in yoga pose classification. ICDAM 2025, London.
• Patel, V. (2024). Machine learning-based hand gesture recognition for immersive gaming.
• Patel, V., & Mehta, K. (2024). Emotion-aware music recommendations: Evaluating custom CNN vs. VGG16 and Inception V3. IEEE FMLDS.