Web Version
Personal Platform with Agentic Technology Reports Pipeline
Mar 2026 – Present
matiasdasilva.com/projects
I designed and developed a personal platform for publishing weekly technology intelligence reports across four domains: Artificial Intelligence, Cloud, Multi-Industry, and Integrator analysis. The project evolved from a static manual-publishing site into a fully automated 9-step agentic pipeline coordinating 7 specialized AI models.
The Pipeline
The system deliberately combines model types based on their strengths: large language models via AWS Bedrock with Mistral Large 3 for high-quality long-form report generation, search-specialized models (Tavily API, GPT-4o) for parallel real-time web research, and lightweight local models (Qwen2.5-Coder via Ollama, MarianMT by Helsinki-NLP) for executive brief synthesis, podcast scripting, and bilingual ES/EN translation with minimal GPU footprint.
The full flow — parallel search, sequential generation of four domain reports, executive brief synthesis, neural translation, and automated deployment to Vercel via Git — completes in under 9 minutes of compute time. Reports are published simultaneously in Spanish and English.
Podcast
Podcast scripts are generated automatically within the pipeline via Ollama/Qwen2.5-Coder. For audio production, Gemini 2.5 Flash Preview TTS was tested as a candidate to fully close the loop, but a direct comparison with NotebookLM revealed a clear gap in conversational naturalness and audio quality. For that reason, audio production is currently done manually with NotebookLM, until Google brings that same quality to its public API.
Tech Stack
AWS Bedrock / Mistral Large 3 · OpenAI GPT-4o · OpenAI GPT-4o-mini · Tavily API · Ollama / Qwen2.5-Coder · MarianMT Helsinki-NLP · NotebookLM · Next.js · Vercel · Git