At FrameFluence, I built retrieval-grounded LLM features with guardrails and evaluation loops, shipping them from prototype to production. I also owned API integration and pipeline debugging to improve response quality and latency.
For the Amazon ML Challenge, I built a multimodal price predictor across more than 75,000 listings, combining text and image signals with BERT, CNN features, and XGBoost. I placed in the top 5%.
I built Interp-n, a satellite frame interpolation engine using RAFT and GRU to turn GOES and INSAT-3D frames into a continuous timeline. On GOES-18, it scored 38.92 PSNR and 0.954 SSIM.
I also built NutriAI, a food calorie estimation pipeline, and ReszVault, a PDF intelligence workspace with grounded chat and citations. In ReszVault, retrieval evaluation increased precision at five from about 61% to about 84%.
