samiel azmaien.

I lead the product and customer discovery. I help companies choose which workflows we will initially support and define success criteria of the test.

As agents operate in different companies, each request involves a different set of permissions. We tie the agent’s proof to the company receiving the request and to the particular request itself. In case the permission runs out of time or is revoked, the next check must deny access even if an earlier request succeeded.

We start building our network by helping two companies conduct one useful interaction. The early discussions made it easier for me to focus on this: find the recipient of the request, give an example based on his real job and decide which requests must succeed and which must fail.

I was selected for Entrepreneurs First's Selection Day. We have paid pilots running and are talking to a16z, Founders Inc. Blueprint, Alliance (ALL19) and Drive Capital.

I co-developed the ISEF research project, designed the predictive caching approach and experiments, and later I was the sole author of the journal article based on that work. I'm also a co-inventor of the patent for the GPU project.

A predictor may end up negating its own advantage if inference cost and memory cost are too high. Thus, I had to take model compression, prediction accuracy, and transfer time as one systems issue rather than three different performance metrics.

A system's contribution is convincing only if the benchmark and simulation assumptions and the artifact set are clearly laid out. Thus, the public repository distinguishes between the publication record and reproducibility package.

The paper claims 85% L1 cache hit rate, 40% reduction in memory cost, and 25% improvement in data transfer rate. The work was presented at the Georgia Science and Engineering Fair where it placed first place for all categories and placed as a finalist in Regeneron ISEF. It was published in IJSCE in January 2026 and resulted in a granted patent.

I developed the tabular features pipeline, trained logistic regression and dense network baselines using two hidden layers, computed metrics and figures, and described the process in the accompanying manuscript.

The public data sample is tiny and does not replicate the bigger experiment performed for the paper. The critical point was distinguishing an exploratory research finding from one that could be perceived as clinical validation.

In medicine, machine learning models require splits that do not depend on the patients, provenance, baselines, confidence intervals, and error analysis before accuracy becomes relevant. A limitation described explicitly is better than a metric with no context.

I published the paper in the International Journal of Advanced Pharmaceutical Sciences and Research in 2024.

I created the Jupyter workflow for data preprocessing, regression models experimentation, sigmoid score calculations, and visualization in rasters, which I documented in a 34-page development logbook.

There was no clean schema or reproducible pathing between election tables, precinct geography information, and files stored on Google Colab. I needed to be careful not to present the score as an estimate.

The cleaner and more professional a map looks, the less care it takes in the methodology. It is just as important to have reproducible inputs, independent labels, train/test split, and clear language as a good model.

The notebooks and full logbook are available publicly. The code is labeled as exploratory and the data/environment/validation needed for a reproducible release are clearly stated in the repository. This research won 1st place at the Georgia State Leadership Conference hosted by TSA.