KALKI 2.0
A local-first AI assistant for the Windows desktop
AI & Software · 2025–2026
Case studyServices — Applied AI
Applied AI and automation.
Practical LLM integrations, automated workflows and local-first AI tools.
AI is useful when it solves a real task. We help teams work out where it earns its place, then build the feature or automation with deliberate choices about data, cost and human oversight.
Our own product shows how we think about this. KALKI is a local-first desktop assistant, published on the Microsoft Store, that sends private queries to a model running on the machine and routes heavier reasoning to cloud providers only when the user chooses. Every system action shows a dry run and waits for confirmation.
The same principles carry into client work: local models where the data is sensitive, cloud models where they make sense, and guarded boundaries around credentials and system actions.
Who it’s for: Businesses reducing repetitive work, founders adding AI features, and teams weighing local or private model options.
What you get
- AI features and automation scoped to a real task
- Prototypes using local models where they fit
- Voice input, transcription and audio interfaces
- Document summaries and information extraction
- Guarded boundaries for credentials and system actions
Typical stack
- Python
- Ollama
- OpenAI API
- Gemini API
- Groq
- VOSK
Related work.
- 01
Questions, answered.
How do you protect business data?
We match the setup to how sensitive the data is. For confidential work we can run local models, for example through Ollama, so data never leaves your hardware. Where cloud models make sense, we add guarded routing and redaction.
Which AI models do you work with?
Local models through Ollama, and cloud APIs including OpenAI, Gemini and Groq. For voice input we also use offline speech recognition such as VOSK.
What kind of work can AI automate for us?
Repetitive tasks with a clear definition of done tend to suit it best, such as document summaries, information extraction and transcription. We scope against a real task first, and will say so if AI is not the right tool.