Srinath Sridhar Email & Phone Number

California, United States

San Francisco, Ca, Us

AI and specifically NLP has made tremendous progress in the early R&D phase over the last 2 years. The progress we will make in their industrial applications over the next decade is truly going to be remarkable. Regie.ai, our recommendation engine, is going to help bring all this R&D effort to help businesses with the content they use to communicate.

PiNK test is the future of early detection of Breast cancer using blood draws instead of mammograms. Using cutting-edge AI to leverage multi-analyte liquid biopsy and clinical data, PiNK test aims to improve specificity and sensitivity simultaneously to over 99% which is significantly higher than mammograms. The liquid biopsy test still costs roughly the same. The project is still in stealth stage, is run by incredible high energy repeat entrepreneurs who have founded multi-billion dollar public companies in the early cancer detection space before, and has professors in computer science and medicine, as well as valley engineers dedicated to making this successful. Happy to put you in touch if you'd like to help bring this to the real world from its early infancy.

Investor in several Silicon Valley startups.

San Francisco, California, Us

Mountain View, California, Us

My work at BloomReach can be summarized as building version 1.0 of many of its products and technologies. Every single one of these products below were built from complete scratch all the way to production. About half were built single-handedly and half with a very small team of 1-2 other engineers.BloomSearch: Author of the first production data pipeline of BloomSearch the flagship product of BloomReach. The pipeline involved several stages that incorporated crawling, processing of raw HTML pages into structured format, indexing into lucene, clustering to determine page similarities and finally a large-scale max-flow-based optimization on the link graph. BloomLift: Single author of the first version of BloomLift (BloomReach's second revenue generating product) targeted to improve the ads landing pages. The work involved analyzing queries, landing pages and computing regression based scores for the best performing pages. Was also later involved in leading a larger team to expand the product into the e-commerce space.BloomSocial: Led a small team of 2 engineers to build the very first version of BloomReach's social product from scratch. The product went through several iterations and live pilots across 3 customers.Zinc: Co-authored Zinc (along with Joshua Levy), the first open-source version control system for data where traditional systems (git, mercurial etc completely breakdown). See: https://github.com/bloomreach/zinc. Zinc is used extensively inside BloomReach to maintain, version and deploy very large data files. Analytics: Single author of the first version of the full BloomReach analytics stack similar to Google Analytics, Coremetrics or Omniture. This involved writing the tracking javascript to fetch a pixel to log user and web-page centric information. Writing BloomReach's first hadoop jobs to process and aggregate these logs on a daily basis. Setting up BloomReach's first databases to support OLAP and display of the analytics dashboard.

First 100 engineers of Facebook. Search Ranking: Built the first machine learning based search ranking algorithm for Facebook to improve on a rules based method. This involved collecting all the search query, click-through and downstream action logs, defining and building features, defining loss functions, running a regression to find coefficients for ranking and finally deploying the new system into production.Query Classifier: Built a query classifier to determine the most likely type of result a user is searching for. This was a Bayesian classifier that used search queries and click-throughs to determine the most probable type given the words of the query.

Mountain View, Ca, Us

Duplicate Image Detection: Developed an algorithm based on locality sensitive hashing to automatically detect duplicate images. The primary purpose of this was to have human raters blacklist images and have automatically all the other near duplicate images blacklisted immediately as well and removed from Google product search. Distinguishing Products/Accessories: Developed an algorithm to automatically distinguish products from accessories so that search results can be appropriately ranked based on whether a user is search for a product like an iPod or an accessory to an iPod. This work involved collecting search and clickthrough logs, defining features like words and price and clustering the products so that accessories can be separated from the products cluster.

Pittsburgh, Pa, Us

Clustering: We developed several fast algorithms to perform clustering. Specifically, one of our algorithms solved the problem of planted bisection on graphs. Bisection or Max-Cut is a well known NP-complete problem. But if the clusters are 'planted' , that is, the vertices of the graph are sampled from two clusters that are separated beyond a certain threshold, then our algorithms run in polynomial time and accurately recompute the clusters. Another algorithm of ours solved a relaxation of Max-Cut where vertices can belong to one of the two partitions fractionally. We used an expectation-maximization heuristic to compute the solution and showed formally again that if the underlying clusters are well separated and the mixing is well-defined, then our algorithm provably finds the optimal solution. Our first algorithm was tested on simulations and in real human genetic variations called SNPs using multiple ethnic groups and our second algorithm was tested on admixed populations. Steiner Minimum Tree: We developed several fast algorithms to find Steiner minimum trees over hypercubes, a well-known NP complete problem. We showed that our algorithms will run in polynomial time if there exists a solution which has cost at most a logarithmic additive factor over the number of dimensions of the cube. This showed that the problem was fixed parameter tractable, an open question until that point. We tested our algorithms on simulated data and on SNPs human genetic variation to reconstruct evolutionary trees.

Oakland, California, Us

Cold Spring Harbor, New York, Us

Phd, Computer Science - Algorithms

Bachelor Of Science - Bs, Computer Science (Honors)