Associate AI Engineer
Best Brains Inc · Feb 2025 – Present
- Building a Socratic AI tutor agent that teaches students by asking leading questions instead of giving direct answers — designed the agent's conversation flow, prompting strategy, and tool/function-calling loop (calculator for step-wise math verification, curriculum content retrieval, answer evaluation, hint generation).
- • Built a RAG pipeline with semantic search, embeddings, and Crawl4AI-based web scraping for AI-ready content ingestion — grounding the tutor in curriculum content, improving factual accuracy and reducing hallucinations in production.
- • Designed domain-specific LLM fine-tuning workflows using LoRA, QLoRA, and post-training quantization to adapt models to the tutoring domain while reducing inference cost.
- • Curated training datasets from image sources using Gemini APIs and Tesseract OCR — selected after evaluating multiple local LLMs — reducing manual dataset-creation effort by ~75%.
- • Integrated YOLO-based object detection to crop question regions with tables, graphs, and diagrams and map them to text — cutting manual cropping workload by ~70%.
- • Automated the end-to-end data-preparation pipeline with Python (watchdog) — reducing manual workload by ~60%.
- • Developed an internal LLM Gateway routing Claude, Gemini, and local LLMs (Ollama) with dynamic model selection, fallback handling, and cost optimization.
- • Implemented LLM observability (latency, token usage, response quality) and refined evaluation workflows — improving downstream performance by ~25%.



