Omar Solano

Omar Solano is an AI engineer and co-developer of Towards AI’s open-source AI tutor, an educational assistant that helps developers navigate technical courses, documentation, and programming problems. His work focuses on making conversational AI systems accurate, measurable, and economical in production.

Based in the Montreal area, Solano studied at École de technologie supérieure before working in artificial intelligence and machine learning. His experience includes fine-tuning diffusion and language models, deploying FastAPI applications, and building retrieval pipelines. He contributed technical writing to Building LLMs for Production and co-authored a 2024 analysis of model distillation and open-source AI, addressing how limited access to proprietary models’ reasoning and predictions complicates knowledge transfer into smaller models.

With Louis-François Bouchard and Samridhi Vaid, Solano developed the Towards AI tutor, combining a LangChain agent, FastAPI backend, and Next.js interface. At AI Engineer World’s Fair 2026, he detailed the system’s architecture and his initial evaluation results:

• Hybrid semantic and keyword search: The tutor combines embedding-based retrieval with BM25, merges and reranks candidate passages, and lets students restrict searches to particular courses or documentation.
• Constrained knowledge-base browsing: A read-only, sandboxed documentation browser limits execution time, output, and commands. Testing found that it added latency without improving retrieval for the student questions evaluated.
• Evaluation grounded in student questions: Solano built an evaluation dataset from actual student questions and staff responses, removed outdated and duplicate examples, and tested retrieval accuracy, answer quality, and multi-turn recall alongside latency and token usage.
• Prompt caching over premature compaction: In his initial experiments, retaining complete conversation histories outperformed sliding windows, summarization, and selective retention. Removing previous tool outputs triggered repeated retrieval, while rewriting history could undermine caching; preserving context improved recall, cost, and latency for the workload tested.