https://nishantnepal.github.io/Building with AI, data, and cloud technologies 2026-02-01T12:08:31-05:00 Nishant Nepal https://nishantnepal.github.io/ Jekyll © 2026 Nishant Nepal /assets/img/favicons/favicon.ico /assets/img/favicons/favicon-96x96.png 2026-01-04T00:00:00-05:00 2026-01-04T00:00:00-05:00 https://nishantnepal.github.io/posts/gen-ai-azure-model-router/ Nishant Nepal

Intro When you are building Gen AI applications where the users are not interacting with the model (backend agentic AI applications or triggering a LLM through api), then you are more or less starting out with a single backend LLM that is the best fit for your call and as your code stabilizes, you upgrade the LLM depending on your evaluation criterias. Standard stuff. But, what happens if you...

2025-12-26T00:00:00-05:00 2025-12-26T12:12:17-05:00 https://nishantnepal.github.io/posts/building-gen-ai-content-safety/ Nishant Nepal

Intro When building generative AI applications—especially customer-facing ones like chatbots—content safety aren’t optional. They’re essential. Large language models (LLMs) are powerful, but they’re also unpredictable. They can generate text, images, and even code in ways that are difficult to fully control. That means it’s not enough for an AI system to simply produce the right answer. It als...

2025-12-09T00:00:00-05:00 2025-12-09T09:46:07-05:00 https://nishantnepal.github.io/posts/data-eng-databricks-mlflow/ Nishant Nepal

This is a quick note to myself for future use cases. Often times, when devloping locally, i tend to point to a local MLFlow instance and then when deploying to upper environments, its redirected to a Databricks hosted MLFlow endpoint. For the most part this is fairly straightforward and databricks has good documentation here. For a quick verification test, here is my sample code import async...

2025-12-03T00:00:00-05:00 2025-12-09T08:46:20-05:00 https://nishantnepal.github.io/posts/building-gen-ai-feedback/ Nishant Nepal

Intro Building generative AI applications introduces unique challenges—challenges we didn’t have to think about when building regular, non-AI software applications (remember those? 🙂). One challenge is feedback - whether it’s a user flagging a response as factually incorrect or an automated agent scanning and triaging issues, feedback is critical because it’s the only mechanism that closes the ...

2025-11-12T00:00:00-05:00 2025-12-15T08:29:33-05:00 https://nishantnepal.github.io/posts/building-gen-ai-evaluation/ Nishant Nepal

Intro My career has fluctuated between software and data as primary focus areas, with infrastructure and DevSecOps as secondary strengths. While there may be more formal definitions of LLM evaluations, the one that resonates most with me is that it’s essentially software testing for AI models. In software testing—unit, integration, performance, etc.—we verify that deterministic code behaves as...