Kaapi
AI Infrastructure for the Development Sector
What is Kaapi?
Kaapi is Project Tech4Dev’s AI platform designed to accelerate responsible AI adoption across the global development ecosystem. Acting as a middleware layer, it enables platforms like Avni, Glific, Evidential, and Dalgo to seamlessly integrate AI functionality—without each having to build or maintain their own infrastructure.
Â
Built as an API-first SaaS platform (launching early 2026), Kaapi will support NGOs with end-to-end AI workflows: conversation management, sector-specific co-pilots, input classification, document handling, and multi-language/multi-modal support—all guided by responsible AI practices.
Platform Architecture
Objective for the Development Sector
Core Features
AI-Assisted Conversation Management
- Guardrails for safe outputs
- Input classification & routing
- Dynamic prompt management
- Retrieval-Augmented Generation (RAG)
- Fine-tuning for contextual accuracy
Input Classification
- Filters out irrelevant inputs (e.g., greetings, gibberish)
- Lightweight, on-premise classification to reduce costs
- Routes messages based on context (e.g., LLM, human review, emergency escalation)
 Language & Document Support
- Multi-language support including Indic & low-resource languages
- Document workflows: PDF (OCR) → Markdown conversion, editing, approval, and knowledge base integration
Multi-Modal Capabilities
- Speech-to-Text (STT) and Text-to-Speech (TTS)
- Image/voice inputs to expand accessibility
Sector-Specific AI Co-Pilots
- Starting with Healthcare, providing contextual support for program teams to aid decision-making and streamline operations.
Evaluation & Benchmarking
-
- Assessment Pipelines: testing, grading, and feedback loops
- Evaluation Framework: integrated with Evidential to assess quality, relevance, safety
- Responsible AI Practices: contextual guardrails, bias checks, privacy safeguards, risk assessments
- Model Ops: versioning, deployment, tracking
- Dashboards: metrics for NGOs to track performance and insights in real-time