Open Source
Models, datasets and tools Daniel publishes openly, with the data, training configuration and evaluation alongside them.
Hebrish
English–Hebrew code-switching speech recognition
English speakers in Israel drop Hebrew words into English all the time: “pick up bamba from the makolet”, “an appointment at Misrad Hapnim”. Speech-to-text models trained on English have never heard those words, so they guess an English-looking one. Stock Parakeet v2 hears “Micollette”; the fine-tune writes “makolet”.
Hebrish is on-device English dictation for English speakers in Israel that writes the Hebrew words correctly, in Latin letters (makolet, teudat zehut). It is a family of fine-tunes of NVIDIA's Parakeet TDT 0.6B v2 and Nemotron streaming models (Jachnun, Sabich and Shakshuka), trained on more than seven hours of code-switched speech from a single speaker, recorded and published since October 2026. It runs offline in the Hebrish apps: an Android voice keyboard plus Ubuntu and Windows tray apps. On the last benchmark, code-switched WER fell from 11.8% to 5.3%, and plain English stayed level. The dataset is open (CC-BY-4.0) and the models are free for non-commercial use, so the method and the results can be checked and reused.
DSR Holdings runs Hebrish as an open-source initiative: the apps are free, with no ads or accounts, and the recordings, training configuration and evaluations are published alongside the models.
Sabich: the main Hebrish model
Hebrish/SabichParakeet TDT 0.6B v2 fine-tuned on code-switching, tech vocabulary and general English. NeMo checkpoint, with sherpa-onnx and GGUF builds for on-device dictation. CC-BY-NC-4.0, access on request.
English–Hebrew Code-Switching Sentences
danielrosehill/English-Hebrew-Codeswitching-SentencesEnglish sentences with everyday Hebrew words and Israeli names in Latin letters, read aloud and recorded on a phone. Term-disjoint train and test splits. CC-BY-4.0.
Sabich v1.0 against stock Parakeet v2
WER is normalised (lowercase, no punctuation). Full per-utterance results, the version history and the training configuration are on the model card; the project is tracked on GitHub.
Browse the dataset
MCP servers
Model Context Protocol servers that give AI agents access to real services, several of them for Israeli platforms that have no official one.
Miklat: public shelters
Public MCP endpoint that lets AI agents direct people to Jerusalem's 198 public bomb shelters.
Agent tooling
Plugins, indexes and building blocks for working with Claude Code and other coding agents.
Claude Code Projects Index
Index of Claude Code repositories and starter templates, from everyday uses to less obvious ones.
Claude Code plugin marketplace
A marketplace of Claude Code plugins: skills, commands and agents, installable by name.
Unofficial MCP Builder
Builds an MCP server from a vendor's API docs, checks it against the live API, and detects when the upstream API drifts.
RTL Email
Sends Hebrew, Arabic, Farsi and Urdu email that actually renders right-to-left in Outlook and Gmail.
More
Everything else Daniel publishes, from datasets and model experiments to MCP servers, Claude Code plugins and agent tooling.