Hi, I'm Nova. I can turn your data into useful and private technology.

Most AI engineers have never shipped inside FERPA, campus networking, and institutional procurement. That is what makes this my rarest skill. Transparency: I invented PATHS Engine. The University of Delaware holds the IP. I surface it here not to represent it as my own product, but because it is the clearest proof of what I can build inside an institution. If your institution is interested in PATHS Engine, reach out and I will connect you with the right people at UD. This kind of engagement works best as an ongoing collaboration, not a single project.

Regulated sectors increasingly want AI that does not ship data to a vendor, and model-agnostic design hedges against lock-in and sudden price changes. The competing pitch, "your data stays in your own cloud tenant," is easier to buy than to self-host, so I have to defend why local beats a private endpoint, and how currently local-model quality has a real ceiling. Best where sovereignty is a hard requirement, and finding the right balance of frontier-AI and local-model quality is a challenge.

When privacy law blocks organizations from pooling data, federated learning is the standard technical answer. The candid read: Task specific ai models are a better fit for most institutions. But they are new and not yet widely adopted. If you are interested in federated learning, I can help you get started with a small pilot.

Clients with private data and limited budgets want a model they own that runs cheaply. The honest part: fine-tuning is frequently the wrong answer, and good prompting plus retrieval on a strong base model often wins. I lead with "do you even need to fine-tune," then deliver if the answer is yes.

"Chat with our documents" is the default ask, and grounding is exactly where most implementations fall down. Basic retrieval is commoditizing fast, so the technology alone will not command a premium. My edge is the grounding discipline and the graph modeling, which clients tend to underrate until a bad answer costs them something.

Many teams have shipped AI features and cannot tell whether they work, so honest measurement stands out against vendors who grade their own homework. The candid read: my own eval tooling is early and the field has funded players. I sell the discipline and the deterministic-grader rigor, and I am upfront that the tooling itself is young.

Good fine-tuning and good evaluation both die without good data, and most organizations have no pipeline for producing it. The honest limit: the value depends on the client supplying subject-matter reviewers, which is usually the real bottleneck, and the output quality at scale is not yet shown. It needs careful framing, because it is the highest-leverage step even when it reads as unglamorous.

Agentic workflows are the dominant theme of the moment, and tooling that manages prompts and skills as versioned assets is underserved. This is also the most crowded space here, every framework is fighting for it, and my agent code is personal-scale rather than a hardened product. I sell the patterns and the discipline, not the specific tools.

Every AI project still needs a working application wrapped around the model, and building that myself means no handoff and one person who understands the whole path. This is table stakes, though. It will not win a bid on its own. It matters as evidence that I can deliver end to end.

Accessibility compliance is a standing budget line in education and government, and AI-assisted accessibility is an underserved niche. Be careful here: the description endpoint is currently a mock and the camera work is an early proof of concept, so the AI parts are designed more than shipped. I sell the client-side engineering and the accessibility framing honestly, without implying a finished product exists.

Domain fluency shortens every engagement and de-risks delivery, and a twenty-year record of shipping is a trust signal that hype cannot fake. The honest framing: experience is a multiplier on the skills above, not a product I can invoice for on its own. It gets me in the room and keeps me there.