Blog • Projects • OSS • About

🇺🇦 #StandWithUkraine

On Feb. 24, 2022 Russia declared an unprovoked war on Ukraine and launched a full-scale invasion. Russia is currently bombing peaceful Ukrainian cities, including schools and hospitals, and attacking civilians who are fleeing conflict zones.

Please support Ukraine by lobbying your governments, protesting peacefully, and donating money to support the people of Ukraine. Below are links to trustworthy organizations that are helping to defend Ukraine in this unprovoked war:

Open Source Reports
April 2026

Rare event - I actually wrote a blog post. Compact Compact Language Detector is the launch post for two new pure-C language-detection libraries: ccld3 and compact-fasttext. Both keep the original weights but drop the C++/protobuf/external-model baggage, and both check 100% against their reference implementations. There’s a WebAssembly demo running CLD2, FastText, and CLD3 side-by-side in the browser.

Another big month for LMPlayground - seven releases shipped to Google Play, adding system prompts, full localization in 28 languages, a device-language-aware model catalog, generation speed metrics, and a RAM-fit gate that blocks oversized models before they crash low-memory phones. The whole release pipeline is now fully automated with fastlane: store listing, “What’s New” notes, and Paparazzi-generated marketing screenshots all roll out from a single push, in all 28 locales.

Two new Spark-related projects shipped their 1.0.0:

Two new sites went live:

  • druk.me as a new personal domain (front page only for now).
  • lmplayground.app as the landing page for LMPlayground (also localized in 28 languages)

LMPlayground

  • 1.4.3 — Added Gemma 4 (E2B, E4B). Per-model generation parameter persistence. Thinking budget control for all thinking model formats via llama.cpp PEG parser (Qwen3, DeepSeek, Gemma 4).
  • 1.4.4 — Generation speed metrics (response time and tok/s per message). Added LFM2.5 350M. Model hint on conversation reload. Automated Google Play deployment via GitHub Actions.
  • 1.4.5 — Performance and stability improvemtns. Hide tok/s and time stats while generation is still in progress.
  • 1.4.6 — Fixed token speed stats not showing for all messages in a session.
  • 1.5.0 — System prompts library with per-model MRU. Localized UI in 28 languages (Arabic, Chinese, Hebrew, Hindi, Japanese, Korean, Polish, Portuguese, Spanish, Thai, Ukrainian, Vietnamese, and more).
  • 1.5.1 — Device-language-aware model catalog — splits models into “Supports your language” and “Other models” using HuggingFace model card language declarations. In-app language picker in Settings via AppCompat per-app locale API.
  • 1.5.2 — RAM-fit gate on model load — refuses weights larger than 70% of total device RAM with a clear error, preventing the crash mmap’d models trigger on low-RAM phones. Performance and stability improvemnts.

spark-cli-skills

  • 1.0.0 — Initial release. Ships the use-spark base skill with a complete command reference for the Spark CLI (mail, calendar, contacts, teams, meetings) plus 13 recipes covering inbox triage by category, morning standup, meeting prep, invitation handling, scheduling, end-of-day review, weekly digest, multi-account review, topic timelines, stakeholder briefs, shared-inbox health, team workload audits, and calendar audits.

spark-claude-extension

  • 1.0.0 — Initial release. Ships Spark.mcpb, a Claude Desktop extension bundle (MCP server) that gives Claude read-only access to Spark on macOS - browse accounts, folders, and emails with Gmail-style filters; search by topic (keyword + semantic); read full threads with headers, bodies, and attachments; list calendar events and find mutual availability; look up contacts and team info; list and read meeting transcripts with summaries and notes. Companion Spark.skill archive bundles the CLI skill for direct import.

ccld3

  • Initial release — Compact Compact Language Detector 3. A pure-C inference engine for Google’s CLD3 model with zero dependencies and weights embedded in source - no C++, no protobuf, no external model file. 109 languages, ~1.4 MB of model data. Validated against the official C++ implementation: 100% agreement on 3000 samples. CMake and SwiftPM build setups.

compact-fasttext

  • Initial release — Compact FastText. A pure-C inference engine for Facebook’s lid.176.ftz FastText model with zero dependencies and weights embedded in source - no FastText runtime, no external .ftz file. 176 languages, ~2.6 MB of model data. Validated against the official Python implementation: 100% agreement on 3000 samples. CMake and SwiftPM build setups.

Writing

  • Compact Compact Language Detector — Long-form launch post for ccld3 and compact-fasttext. Covers why CLD2 worked for Spark, why CLD3 (protobuf) and FastText (separate model file) stopped feeling compact, and how repackaging the original Google/Facebook weights as pure-C inference engines brings them back to “drop in and forget” territory. Includes a side-by-side comparison table and a WebAssembly demo running CLD2, FastText, and CLD3 in the browser.

Websites

  • lmplayground.app — New marketing site for LMPlayground. Sections cover on-device privacy (zero packets during inference), 25+ curated mobile-friendly models, the three-tap install flow, and the 28-language UI. The site itself is also localized in 28 languages.
  • druk.me — New personal domain. Single-page card site for now, deployed via GitHub Pages with a custom domain.
OSS Report April 2026 Apr 30, 2026 OSS Report March 2026 Mar 31, 2026