Full courses. Real structure. Generated in under 60 seconds from any topic. Kyren is an AI-powered learning platform that transforms any subject into a complete, structured, multi-module course with lessons, quizzes, and progress tracking — entirely generated by LLM pipelines, not curated by hand.

The platform uses a multi-stage LLM pipeline: topic decomposition → curriculum generation → per-module content synthesis → assessment generation. Each pipeline stage is independently streamed to the UI, giving users instant feedback and a perceived sub-10s start time even for complex topics.

A production-grade, developer-first web search API built for LangGraph, LangChain, RAG pipelines, and autonomous AI agents. SearchMind delivers clean, structured, LLM-optimized search results via a simple REST API and pip-installable Python SDK — designed as a direct Tavily alternative with superior content extraction and native agentic tooling.

SearchMind's core pipeline runs: Brave Search fetch → domain filtering → safety screening → parallel content extraction (trafilatura + readability fallback) → credibility-weighted relevance ranking → optional Claude-powered answer synthesis. The Python SDK ships with native LangGraph and LangChain tool wrappers, making it a one-line drop-in for any agentic AI stack.

The only on-device TTS platform combining enterprise-grade speech quality, 31-language support, expressive speech controls, and GDPR-compliant zero-data-retention — accessible via a SaaS dashboard, REST API, and Chrome extension. Powered by Supertonic's ONNX inference engine running entirely in-browser via WebGPU.

ReadAloud solves the core privacy failure of every cloud TTS platform: user text is never transmitted. All inference runs in-browser via ONNX Runtime + WebGPU acceleration. The platform ships an OpenAI-compatible /v1/audio/speech endpoint, making it a zero-code migration for any existing TTS integration. An expression tag system (<laugh>, <breath>, <emphasis>, <pause>) delivers natural prosody that cloud alternatives cannot match on-device.

Multi-agent web research with browser automation, visual delta analysis, and synthesized intelligence. ARIA dispatches autonomous sub-agents to explore the web in parallel, compares visual states across time, and produces structured research reports via Groq LLaMA 3.3.

ARIA's architecture decomposes a research query into sub-queries, dispatches independent Playwright browser agents for each, runs visual diff analysis across page snapshots, and synthesizes all findings through a structured LLM reasoning chain. The result is a report that cites live web sources rather than training-data knowledge.

A real-time financial market simulation where AI agents trade, reason, and compete using LLMs and RAG. Thousands of autonomous agent decisions execute per tick, with live market state streamed to connected clients via WebSocket.

Each simulated agent in StockMind holds a portfolio, receives a market state context window, queries a Qdrant vector store for historical analogs, and reasons through a Groq-hosted LLaMA model to determine its next action. The aggregate of thousands of such micro-decisions produces emergent price discovery, volatility clustering, and boom-bust cycles observable in real-time on the dashboard.

Autonomous PR risk scoring and vulnerability detection before every merge. CodeGuard integrates with GitHub Webhooks, analyzes every pull request for security vulnerabilities via GPT-4o, and surfaces risk scores and fix suggestions directly in the PR timeline using real-time Socket.io updates.

CodeGuard's Policy-as-Prompt system allows teams to define custom security rules in .codeguard.yml, which are dynamically injected into the GPT-4o analysis prompt at review time. The Cross-File Taint Analysis engine builds a data flow graph across all changed files using XYFlow, identifying vulnerabilities that span module boundaries and cannot be caught by per-file linters.

AI-driven red-team orchestration platform that automates adversarial security testing workflows. RedForge models attacker reasoning chains, generates targeted exploit scenarios, and orchestrates autonomous scanning and reporting pipelines.

A complete toy programming language built from scratch, rooted in Gujarati syntax. KemLang ships with a full interpreter, CLI, React-based web playground, and a Python/FastAPI backend — demonstrating end-to-end language tooling implementation as a culturally grounded systems project.