ARNAB KAR January *, ****

adjlk8@r.postjobfree.com Unit 145, 4557 11th Avenue NE,

1-984-***-**** Seattle, Washington 98105.

arnabkar.github.io United States.

INDUSTRY EXPERIENCE

• Senior Data Scientist, Sirion Labs (AI-assisted Contract Management), 2020-now. EDUCATION

• Ph.D. student (affiliated with a lab), M.S. (May 2020), Duke University, 2018-2020.

• Dual degree: B.S. and M.S. (Computer Science), IIIT Allahabad, 2013-2018. TECHNICAL SKILLS

• C++, Python, Shell, TensorFlow, SQL, PostgreSQL, MongoDB, Docker, Kubernetes, Redis, Go, Java, Design-patterns.

EXPERIENCE (selected)

• Language-model assisted OCR correction / Automated layout-detection services (Senior Data Scien- tist)

(Manager: Anu Engineer, CTO, Sirion Labs), September 2020 - now.

Put together a post-OCR (Optical Character Recognition) mistake correction service. It takes the OCR generated output, detects the unusual/malformed words, and corrects them. This is ac- complished employing a hybrid approach of language-model predictions combined with word- distance measures. This service was containerized, productionized, and offered up as a REST-API service (also interactivity provided via Pub-Sub and Cloud Storage).

Leading an automated (contract) layout detection service – to help analyze contracts from a graph theoretic perspective.

• Influence function for Inference (Pre-doctoral internship: Distributed Systems / Compressed Sensing)

(Guide: Prof. Dan Alistarh, Institute of Science and Technology – Austria, February 2018 - August 2018)

In this project we found, from a dataset, a core-set that can effectively and efficiently summarize a dataset. This we accomplished by using (approximate) Influence Function – a non-parametric statistics tool, which quantifies the influence of each data-point on other data-points (for a base- model). Information theoretic and compressed sensing tools were used in this project. Influence Functions for Inference, with Prof. Dan Alistarh (to be submitted in 2021 calls – NeurIPS/AISTATS)

• Multi-agent Generative Adversarial Networks (Student Research Associate: Deep Learning / Game The- ory / Information Theory)

(Guide: Prof. Vinay Namboodiri, Indian Institute of Technology – Kanpur, India, May 2017- July 2017)

In this project, we designed a deep neural network architecture which could approximate a proba- bility distribution. The net’s objective was to learn the semantic information provided in the data samples and to generate an image sample by blending together independently generated fore- ground and background images. The designed architectures and objective functions had information- theoretic and game-theoretic foundations.

(Paper in preparation) Synthesizing images by entangling foreground-background pair to address mode- collapse, with Prof. Vinay Namboodiri

• Fixing Failures in a Component Based Design setting (Summer Research Project: Formal Verification)

(Guide: Prof. Ansuman Banerjee, Indian Statistical Institute – Kolkata, India, May 2016 - July 2016)

With this project, we tried finding out if set of (software) components can achieve a Global Objec- tive. If not, we identified the smallest fix (in some sense).

In another study, we simulated a few attack models on the branch predictors of x-86 architecture. The consequences of the attacks were studied.

Branch Predictor Attacks : An Empirical Exploration: HiPC - 2016. (poster), CCSN - 2016. FixCD : Fixing Failures in Component Based Designs: DATE-2018 review, resubmitted VMCAI-2020. PROJECTS (selected)

Designing Compiler for relational database hardware accelerator to minimize query-execution time and hardware communication [readying for publication].

Distributed training: Resnet-50 on 2048 machines in under 14 minutes (world record at that time).

Increase your coverage: Analysis of how Google-Playstore apps recommender system works (by analyzing the graph constructed by crawling Google-Playstore). [heading for publication]

Citation leeches: Designed algorithms and heuristics to discover self-reinforcing malicious clusters

– of agents trying to boost each other’s citation counts – in an academic citation network.

Locality of search and computation: Designed/implemented a peer-2-peer file storage system as course project. We tried to use locality sensitive hashing (LSH) and PASTRY protocol to come up with a way for location sensitive traffic data routing, which increased through-put, minimized latency, and increased availability.

Ensuring a changing distribution: designed oracles that provides data samples (from a distribu- tion) through time-steps such that the samples are mutually independent (to verify the robustness of active-learning and online learning algorithms) (Duke, CS 634: Geometric Algorithms).

Mechanism design for preventing undercutting by incentivized consensus protocols for Blockchain technologies (Duke, CS 590.2: Computational Microeconomics). RELEVANT COURSE-WORK

Duke University

Geometric Algorithms, Computational Microeconomics, Privacy and Fairness, Advanced Linear Al- gebra, Design and Analysis of Algorithms, Probability and Measure Theory, Stochastic Processes (au- dited), Datacenter Architecture, Coding Theory and Data Science. IIIT Allahabad (representative)

Distributed Systems, Graph Theory, Embedded Systems and their verification. MOOCs (Coursera, and others.)

Machine Learning, Game Theory Online, Neural Networks for Machine Learning, Computer and Infor- mation Security (Duke).

SERVICE

Machine Learning - STA671: TAed for Prof. Cynthia Rudin, Duke University (Spr. 2019: duties)

Design and Analysis of Algorithms: TAed for Prof. Amit K. Dhar, IIIT Allahabad (Spring 2018).

Probabilistic Graphical Models: Organized for Prof. Gaurav Sharma lab. sessions for his Govern- ment of India (GIAN) course. (2017 winter break)

Theory of Computation: TAed for Prof. Amit K. Dhar, IIIT Allahabad (Fall 2017). AWARDS / ACHIEVEMENTS

Duke University Ph.D. Fellowship.

Institute gold medal for first rank in the graduating batch of 2018 (undergraduate).

Undergraduate Merit Incentive Award in 2013, 2014 and 2016 (awarded to top 5 out of batch of 289).

OTHER PUBLICATIONS

Checking for infinite explosion in Chip-Firing game in O(n2

), with Amir Goharshady.

Contact this candidate