Machine Learning Engineer building intelligent systems for real-world problems.
Applied ML · Computer Vision · Statistical Modeling · AI Engineering
Explore my projects through the technologies and ideas that connect them.
Real independent projects, presented with clear claim boundaries. Interactive pieces are browser adaptations; reported results remain tied to their original experiments.
A JAX-native environment for reinforcement learning: vectorized resets and steps, partial observations, action masks, massive batches, GPU benchmarks, video and Streamlit inspection.
A five-package CRISP-DM monorepo for ventilator pressure prediction: stacked LSTM, group-aware evaluation, Docker pipeline, FastAPI inference and CI/CD releases.
Attention U-Net over 3,064 MRI/mask pairs. A synthetic RGB input combines CLAHE, Sobel and DoG; the final test run reports F1 0.763, precision 0.859 and recall 0.779.
A CPU-reproducible study of how recurrent spiking networks transform grapheme events into semantic population structure, with ablations and a paired ANN baseline.
Two project snapshots showing both sides of applied ML: a model that passes an experimental evaluation, and a deployment gate that correctly blocks an inadequate prototype.
The tracked 200-breath, one-epoch pipeline report records RMSE 14.75 and group gaps across resistance/compliance settings. The automated business evaluation rejects production deployment and flags strong bias.
The final notebook evaluation uses a 512×512 Attention U-Net with synthetic RGB channels: CLAHE for local contrast, Sobel for edges and Difference of Gaussians for intermediate structure.
The strongest project story is not a bigger number. It is a reproducible experiment with controls, ablations and explicit limits.
Ordered processing is required to construct the representation, while final semantic organization is expressed mainly as a sparse population/rate code—not precise spike timing.
The study supports supervised semantic organization. It does not establish unsupervised language acquisition, energy efficiency, reasoning or SNN superiority over ANNs.
I approach Machine Learning as an engineering discipline: understand the decision first, build the smallest useful system, and evaluate what can go wrong.
Define the decision, constraints and cost of error.
Create a baseline and choose honest evaluation signals.
Develop reproducible data and model pipelines.
Stress-test assumptions, edge cases and failure modes.
Integrate, observe and improve the whole system.
I work at the intersection of machine learning, software engineering and applied problem solving. I’m especially interested in systems where a model has to survive contact with imperfect data and real operational constraints.
Professional work stays confidential. Independent repositories make the engineering process and evidence inspectable.
Predictive modeling · Experiment design · Time series · Evaluation
Image processing · Segmentation · Detection · Measurement
Python · TypeScript · APIs · Data pipelines · System design
Reproducible experiments · Ablations · Baselines · Claim boundaries