Jonathan Ve Vance Email & Phone Number

Bengaluru, Karnataka, India

- Owner of two online BERT based query-ad ranking models ensuring quality of Microsoft ads, from training to deployment. - Worked on reducing the latency of online models using efficient attention models like Longformer/LongT5, and achieved significant reduction in latency at negligible drop in performance metrics.- Owned the end to end strategy on detecting sensitive queries, from multi-stage knowledge distillation training to online deployment. - Worked on inducing geographical (or world) understanding to online models and the stack, helping to reduce customer complaints from travel advertisers.

Bengaluru, Karnataka, India

* Designed a graph based algorithm to automate discovering insights from tabular data and creating a "data story", supported by charts and compelling insights, based on the columns of interest. * Conducted a large scale user study where users were asked to compare our data stories v/s those generated using prior art. Results suggest that our data stories are 30% more insightful and 25% more informative.* Novelty of algorithm ensures higher exploration of the graph space, which granted us a patent (USPTO).

* Designed an efficient semi-supervised convolutional network based framework for anomaly detection in industrial products moving on conveyer belts, captured by stationary cameras in controlled environments.* Achieved F1 score of 0.85 and reduced false positives by 60% over baseline autoencoder models. Deployed the model on an in-house edge device at 80 FPS using OnnxRuntime / Intel Openvino toolkit.* Used 3D convolution (C3D) and multi-instance learning (MIL) for video anomaly detection in CCTV footage and achieved 0.75 F1 score on in-the-wild scenarios ranging fights, stampedes, traffic violations, etc.

Pune, Maharashtra, India

* Designed a framework to utilize the time series information from the past several hours in an assembly line to predict the number of industrial parts that will be produced in the next 1 hour.* Dealt with the small dataset size and the lack of internal measurements of machines such as temperature, revolutions per minute, etc using extensive domain-based feature engineering - to quantify suboptimal waiting periods between machines in the assembly line, temporary machine slowdowns, etc.* Posed the problem as an autoregressive time series formulation and reduced RMSE by 20% using Random Forests model over the baseline model.

Chennai Area, India

* Proposed that the relationship between UTS and friction stir welding process parameters should be linear.* Achieved average absolute relative error of less than 5%, and an error reduction of 40% compared to more complex baseline models like neural networks, which were recommended by prior literature.* Second author of the paper published in Journal of Mechanical and Energy Engineering (JMEE) 2019.

Master Of Technology - Mtech, Data Science

Bachelor Of Technology - Btech, Engineering Design