At Impact Analytics, I build and productionize AI systems for retail, including a unified time-series foundation model for demand forecasting across products and clients.
I adapted and fine-tuned Moirai for enterprise demand data, reducing weekly WAPE by 15%, then added store-level representations and gated cross-attention adapters for a further 5% reduction. I designed the end-to-end pipeline from preprocessing and tokenization through training, inference, and evaluation.
I also compressed a Qwen-family retail catalog model, reducing parameters by 25%, model-weight memory by 63%, and end-to-end latency by 30% while maintaining generation quality. My work includes LoRA fine-tuning and vLLM serving optimization across batching, quantization, prefix caching, and chunked prefill.
Previously at Bidgely, I built demand forecasting and intervention pipelines that helped reduce customer billing costs by 6%. My earlier work at Texas Instruments and IIT Madras grounded me in analog IC design, CTDSMs, silicon validation, and research-driven engineering.
