Senior .NET developer transitioning into AI engineering. Building real AI applications — RAG pipelines, LLM integrations, AI agents, Image to Text extraction — using Semantic Kernel, OpenAI, and .NET 8, deployed on Azure. Everything built in public.
// projects
Every line of code is open source. Try the live apps, watch the recorded demo, or read the code.
First .NET + OpenAI integration. Console app with full conversation history, error handling, and secure API key management via user secrets.
Real-time token streaming like ChatGPT. Uses IAsyncEnumerable and await foreach to print each word as it's generated — zero waiting for full response.
Production-grade AI chatbot exposed as a REST API. Built with Microsoft Semantic Kernel. Supports multi-session conversation history and is deployed on Azure App Service.
Upload any PDF and ask questions about it. Uses RAG — Retrieval Augmented Generation — to find relevant chunks via vector search and answer using only your document's content.
An autonomous AI agent with 13 tools across 4 plugins. Uses the ReAct pattern via FunctionChoiceBehavior.Auto() — decides which tools to call, calls them, reads results, and reasons to a final answer. All automatically.
Upload any image and let GPT-4o analyse it. Extracts visible text (OCR), answers questions about the image, performs full scene analysis, and pulls structured data from invoices and receipts — returned as typed JSON ready for any system.
Speak and the AI speaks back. A full three-stage voice pipeline: Whisper transcribes your audio to text, GPT-4o-mini generates a concise conversational answer, and OpenAI TTS converts it to natural speech — returned as audio you can play instantly in the browser.
Three AI agents reason it out live, not behind a spinner. A Researcher gathers the angles, a Critic pressure-tests them, and a Synthesizer merges both into a final answer — each step streamed to the browser the instant it's ready over SignalR. Built to solve the actual blocker enterprises hit with autonomous agents: seeing how a decision was reached, not just the decision itself.
Push to talk and watch the answer stream back in real time. Free browser-native speech-to-text feeds GPT-4o-mini, which streams its reply token-by-token over SignalR — no per-minute transcription cost, no waiting for the full response, and the conversation remembers context until you reset it.
An AI that answers with the Wi-Fi turned off. Phi-3-mini runs on a laptop CPU through ONNX Runtime and Semantic Kernel, so there's no API key, no cloud and no per-token cost. The honest trade-off: about 3 tokens/s, and a small model that can get facts wrong.
// tech stack
The stack I'm building with and learning deeply.
// learning journey
12 months. Building in public. Every week a new concept, every two weeks a new app.