Blog Jul 24, 2026

Inside SwiftF0: Architecture, Decoding, and Practical Capabilities

SwiftF0 is a compact neural fundamental-frequency estimator designed for monophonic audio. Its interesting idea is not simply that it is “small” or “fast.” The model reorganizes the pitch-estimation problem before applying a neural network: it converts waveform audio into a deliberately narrow spectral representation, learns local harmonic patterns with five convolutional layers, projects them onto a logarithmic pitch axis, and then decodes a continuous frequency from a local probability distribution. This article follows that path end to end—from samples to pitch_hz and confidence—and examines what those outputs can and cannot tell us.

Blog Nov 5, 2024

Sequential Chains with Memory using LangChain

The latest LLM’s provide much better performance than previous approaches at the language modelling task when approached as a one-shot task. Many a times, we want to chain together sequences of such LLM’s, prompting each other to create more complicated pipelines compared to simple one-shot prompts. We might also want to process the LLM outputs in order to return structured objects like JSON or XML for tasks like classification or text summarization.

Blog Nov 4, 2024

Exploring Vercel AI SDK and next.JS

2024 has been the year of LLM’s and other AI models with powerful generative capabilities across the domains of language, image, video and audio. While foundation models provide the essential building block for creating intelligent features in our applications, building AI-native user experiences on our devices requires a new way of thinking about how we integrate these AI models into our everyday software applications. Vercel, the creators of Next.JS are one such player in this market, easing up the deployment process for the next generation of applications creating native AI experiences on the web.