I built SWALPA (https://swalpa.org) because learning spoken Kannada is a New Year's resolution for me and I realized most language apps teach you formal grammar, but not how to negotiate with an auto-rickshaw driver or handle the "daily chaos" of the city.
THE AGENTIC STACK (BUILT WITH ANTIGRAVITY)
I wanted to see how far I could push a "Human-in-the-loop" AI workflow to create high-quality content in a low-resource language.
* The Coder: Antigravity (an agentic coding AI) handled the heavy lifting—writing the JS game engines and drafting the core lessons based on a custom "grammar spec."
* The Knowledge Base: I used Gemini Deep Research to build a high-fidelity reference of Kannada linguistics. This acted as the "Source of Truth" to prevent AI hallucinations.
* The Content Engine: I fed my core curated lessons into NotebookLM to generate supplemental podcasts, slides, and flashcards to boost recall.
* Architecture: Simple and Boring—Static site (MkDocs Material) + Vanilla JS + Firebase.
THE EXPERIENCE
Instead of "The cat is on the table," SWALPA uses interactive games to simulate real-world Bangalore scenarios:
* The Auto-Rickshaw Negotiator: Use the right phrases under timer pressure to get the driver to use the meter ("Meter Haaki!").
* Rapid Translation: A high-pressure mode where you hear a Kannada phrase and must pick the correct translation from four options under a timer.
* Agglutination Mastery: Interactive drills for Kannada's unique word-stacking grammar.
The platform includes 10 core text lessons and Duolingo-style gamification—including badges, streaks, and an activity heatmap.
I am a Software Engineer at Google, but this is a personal passion project. I would love to hear your thoughts on the AI workflow or the "street-smart" approach to language learning!
maurys•1h ago
I have added two deeper "Behind the Scenes" pages for those interested in the linguistic philosophy or the specific agentic workflow:
This section dives into the "Survival Kannada" philosophy. It explains why we prioritize high-frequency scenarios like negotiating auto fares or understanding local etymology (e.g., why tech hubs like Marathahalli still carry the "-halli" village suffix) over standard textbook grammar.
This page breaks down the "Local-First" stack and the agentic loop:
* The Staging "Brain": I maintain a private repository where I use Gemini Deep Research to crunch raw linguistic data. This verified "Spec" is what I then feed to Antigravity (the coding agent) to build out the game engines and lessons.
* Portability: The framework is designed to be "forkable." A sister project for Tamil (konjam.org) is already being built using these same architectural principles.
I am happy to dive into the weeds on the prompt engineering, the TTS pipeline, or the Bangalore-specific nuances if anyone is interested!
maurys•1h ago
I built SWALPA (https://swalpa.org) because learning spoken Kannada is a New Year's resolution for me and I realized most language apps teach you formal grammar, but not how to negotiate with an auto-rickshaw driver or handle the "daily chaos" of the city.
THE AGENTIC STACK (BUILT WITH ANTIGRAVITY)
I wanted to see how far I could push a "Human-in-the-loop" AI workflow to create high-quality content in a low-resource language.
* The Coder: Antigravity (an agentic coding AI) handled the heavy lifting—writing the JS game engines and drafting the core lessons based on a custom "grammar spec."
* The Knowledge Base: I used Gemini Deep Research to build a high-fidelity reference of Kannada linguistics. This acted as the "Source of Truth" to prevent AI hallucinations.
* The Content Engine: I fed my core curated lessons into NotebookLM to generate supplemental podcasts, slides, and flashcards to boost recall.
* Architecture: Simple and Boring—Static site (MkDocs Material) + Vanilla JS + Firebase.
THE EXPERIENCE
Instead of "The cat is on the table," SWALPA uses interactive games to simulate real-world Bangalore scenarios:
* The Auto-Rickshaw Negotiator: Use the right phrases under timer pressure to get the driver to use the meter ("Meter Haaki!").
* Rapid Translation: A high-pressure mode where you hear a Kannada phrase and must pick the correct translation from four options under a timer.
* Agglutination Mastery: Interactive drills for Kannada's unique word-stacking grammar.
The platform includes 10 core text lessons and Duolingo-style gamification—including badges, streaks, and an activity heatmap.
Games Link: https://swalpa.org/games
Source: https://github.com/saurabh-net/swalpa
I am a Software Engineer at Google, but this is a personal passion project. I would love to hear your thoughts on the AI workflow or the "street-smart" approach to language learning!
maurys•1h ago
WHY SWALPA? (https://swalpa.org/behind_swalpa/why/)
This section dives into the "Survival Kannada" philosophy. It explains why we prioritize high-frequency scenarios like negotiating auto fares or understanding local etymology (e.g., why tech hubs like Marathahalli still carry the "-halli" village suffix) over standard textbook grammar.
AI ARCHITECTURE & WORKFLOWS (https://swalpa.org/behind_swalpa/architecture/)
This page breaks down the "Local-First" stack and the agentic loop:
* The Staging "Brain": I maintain a private repository where I use Gemini Deep Research to crunch raw linguistic data. This verified "Spec" is what I then feed to Antigravity (the coding agent) to build out the game engines and lessons.
* Portability: The framework is designed to be "forkable." A sister project for Tamil (konjam.org) is already being built using these same architectural principles.
I am happy to dive into the weeds on the prompt engineering, the TTS pipeline, or the Bangalore-specific nuances if anyone is interested!