Pulkit Tandon — Product Engineer

I’m a researcher and engineer who builds scalable, data-efficient ML systems and owns the full lifecycle, from problem formulation and research to deployment. My work includes training-data selection and curation, ranking and recommender systems, and perceptual evaluation and compression. I focus on rigorous evaluation and measurable impact in real-world settings. At Granica, I led the training-data selection initiative end-to-end, including problem formulation, algorithms, scalable scoring and selection pipelines (Ray/Spark + PyTorch), and enterprise pilots on multi-petabyte and billion-row corpora. This work delivered 2–5% targeted metric lift via online A/B testing while keeping adoption friction low through minimal infrastructure changes. This line of work was recognized with an ICLR 2024 Oral and Best Paper Honorable Mention. Previously at Netflix Encoding Technologies, I created CAMBI, a perceptual metric for banding artifacts that was open-sourced in VMAF and adopted into Alliance for Open Media common test conditions, helping influence how next-generation codecs are evaluated. During my PhD at Stanford, I worked on generative and semantic compression (Txt2Vid) and hardware-aware data compression for neural recording (compressive ADCs). I also teach as a Stanford Adjunct Lecturer and co-designed EE274, Data Compression. I’m currently most excited about efficient, scalable ML models and systems, and about translating applied research into practical impact at scale. This includes dataset optimization, post-training methods, curriculum learning, evaluation harnesses, and efficient training. I enjoy collaborating with teams that build ML systems where better data beats more data.

Stackforce AI infers this person is a Machine Learning Engineer with a focus on data optimization and scalable systems.

Experience: 6 yrs 1 mo