At Coherent Market Insights, I designed a trend-discovery engine for CoherentConnect, organizing news into a four-level trend hierarchy. I also engineered a LangGraph buyer-matching layer that surfaces potential buyers from daily signals.
For a newsletter engine, I fine-tuned Qwen2.5-7B-Instruct with QLoRA to replace a paid API. The model beat its teacher on 15 of 20 held-out prompts while producing valid JSON.
I developed a local CRM intelligence engine combining hybrid RAG, lead scoring, and a Qwen-based router to help sales reps prioritize leads and query data in plain English.
At SparkAI.tech, I architected an ETL pipeline for energy datasets and trained Temporal Fusion Transformer models for consumption forecasting. I also built TinySQL, a 125M-parameter text-to-SQL model from scratch, with a schema-constrained decoder for valid SQL generation.

