At MAssist (Inthink Technology), I trained a Qwen2.5-Coder-1.5B Text-to-ClickHouse SQL model with QLoRA supervised fine-tuning and GRPO reinforcement learning. Test execution accuracy rose from 0% to 82.4% using an execution-based reward.
I designed the ML ENGINEER AGENT tree-search experiment protocol and used it to run 228 experiments. This work included running Qwen3-VL-235B within a 7.5 GB memory limit and speeding up DeepSeek-OCR processing by 8.3× with CUDA graphs and a custom Triton kernel.
Previously, as Founding Engineer at Indus AI, I developed and deployed a multilingual text-to-speech system supporting 12+ languages and 20+ speakers. I replaced third-party speech APIs with in-house models, reducing costs by 85%+, and developed a shop classification system with 90%+ accuracy across 9M+ data points.
