How models can adapt at inference time, what is learned during that process, and which trade-offs make adaptation reliable.
Serendeep Rudraraju / M.Sc. computer science
Test-time adaptation, interpretability, and long-context models.
I study how models learn, represent tasks, and use context through controlled experiments, reproducible software, and technical writing.
Selected work
Software and technical tools.
Selected applications, tools, and contributions. Each entry links to its source repository and, where available, releases and automated checks.
Index / 01local audio tooling
cratemind
A local workflow for downloading Spotify playlists, analyzing audio for BPM, Camelot key, and genre, and organizing tracks into configurable DJ-library folders.Python · spotdl · ffmpeg · local audio signals
Index / 02offline-first learning tool
Marginalia
An Android tablet app for handwritten notes beside PDFs, with flashcards, spaced review, and search that reads your handwriting.Kotlin · Jetpack Compose · Room · PDFium · Android Ink
Index / 03developer tooling
latch
A local desktop secret broker: a coding agent asks latch to run a project command with a named secret, and a popup shows the exact command for one-time approval, so the value never enters the conversation or a .env file.Rust · Tauri 2 · React · SQLite · coding agents
Technical writing
Notes on systems, methods, and implementation.
I write explanatory essays and implementation notes about the systems I build and study.
July 2026June 2026June 2026June 2026
The Math Remembers: How to Catch a Model Lying About Being Original
Model-merge forensics using weight arithmetic to test claims of originality in supposedly homegrown AI systems.
Bounded Wrongness: A Field Guide to Probabilistic Data Structures
Bloom filters and related data structures, with attention to one-sided error, tunable false positives, and memory trade-offs.
The Rest of the Crate: Tempo, Key, and Never Stranding a File
Implementation notes for cratemind: tempo estimation, Camelot key detection, downloader reliability, and file organization.
Your Music Has No Genre: Reading It Off the Waveform
A case for audio-derived genre detection when playlist metadata is incomplete, unavailable, or too coarse.
Independent study
Research questions I am currently working through.
My current process is literature mapping, technical writing, and small implementations. These are open study topics, separate from the published work above.
Methods for understanding internal model behavior, connecting mechanisms to outputs, and separating useful evidence from plausible stories.
Related writing: The Math Remembers: model-merge forensicsHiPPO, S4, Mamba-style models, polynomial projection, and the gap between long-context storage and long-context prediction.
Related writing: Memory as Polynomial Projection