What is happening here ?
Sal is an accomplished engineer and researcher. A continuous innovator with a thirst for implementing new creative ideas, he was part of the initial engineering team and spent 3-years (’14-’17) building mission critical software for a ride-sharing startup in Chandigarh and then one of the largest SE-Asian food-tech startups based in Mumbai. After joining grad school(’17) and having been introduced to dedicated research at a top industrial compuer vision research lab in Santa Clara(’18) he graduated with a Masters degree specializing in AI(’19). His Masters supervised research dealt with Depth-wise Adaptive Layers with Conditional Input for Adversarial Training of Unsupervised Segmentation of Ischemic Stroke Lesions: Where he created models for adversarial training of domain adaptation between MRI scans of Brain Lesions. The model performs unsupervised image segmentation from source domain to target domain. This was under the guidance of Dr. Mincheng Gao @UB.
He is currently a full-time doctoral candidate on an F1-Visa at the State University of New York, Buffalo under the guidance of Dr. Govindaraju; whose lab cedar/cubs@UB has produced stellar work in the areas of document analysis and human biometrics including but not limited to publications in CVPR, TPAMI, ICML, ICFHR, ICDAR, ICPR, FG etc.
Sal’s current research interest focuses on teaching machines to reason. This includes learning robust representations for graph based data- a perfect test bed for which is learning mathematical expressions. Other applications include video summarization, expression detection, fingerprint matching, etc other graph matching-ranking problems.
Evaluation Scale :
- 0-30% : Academic Projects, Hackathons, Personal Hobbys
- 30-60% : Mid-sized/Exploratory Projects at Work (MVP, POC), Low Scale <=5k i/ops deployment
- 60-90% : Production Level projects, Published Research, Mid to High Scale >= 20k i/ops deployments
- 90-100% : 404 Not Found
Machine Learning / AI
Deep Learning Frameworks
Machine Learning Libraries
Web Development Frameworks
CUBS/CEDAR, University at Buffalo, NY
- • Research with building a deep-metric learning based Math Information Retrieval System.
- • Graph based representational learning of math expressions in the wild.
- • Relational Graph Convolution to learn mathematical semantics and visual layout.
- • Equation Attention Relationship Network (EARN) : A Geometric Deep Metric Framework for Learning Similar Math Expression Embedding : Saleem Ahmed, Kenny Davila, Srirangaraj Setlurand Venu Govindaraju
- • Paper ACCEPTED at ICPR ’20
- • Research in Action Unit based recognition of Emotions.
- • Explore relational bias between different AU’s
- • Learning Proxy graph based representation for spatio-temporal changes in facial emotion.
- • Paper UNDER SUBMISSION
- • Research with building a feature extractor from images for word representations.
- • Part of an end to end model for visual token recognition in images.
- • Different model architectures experiments includes strided, inception based, residual, dense residual models etc.
- • Summarizing Lecture Videos by Key Handwritten Content Regions: B.U.Kota, Saleem Ahmed , Alexander Stone, Kenny Davila, Srirangaraj Setlur and Venu Govindaraju
- • Paper ACCEPTED at Icdar’19 Workshop : Best Student Paper CBDAR ’19
Masters Supervised Research
under Dr Mincheng Gao - University at Buffalo, NY
- • Adversarial training mechanisms for adapting a source domain to a target domain
- • Adapting all layers of the base model and also adding a conditional input
- • Incremental generative training for closing the gap between predictions of brain lesion segments from different MRI machines.
- • Evaluated using two MR databases of subjects with traumatic brain injuries, acquired using different scanners and imaging protocols.
- • Depth-wise Adaptive Layers with Conditional Input for Adversarial Training of Unsupervised Segmentation of Ischemic Stroke Lesions : Saleem Ahmed, Mincheng Gao
under Dr Andrew Olewnik - Experiential Learning Department, University at Buffalo, NY
- • Developed Ontology automation tools.
- • Used neural nets and random forest for automatic generation of ontology’s for manufacturing production industries.
- • Software developed in parts with CUBRC
Computer Vision Research Internship (signed NDA)
Ericsson, Santa Clara, CA
- • Professional Research Experience: Internship structured around self identification of problem, proposal and experimental solution generation
- • Published to internal Ericsson Journals
- • Increasing Throughput of Image Segmentation Models: Augmentation of Instance segmentation models,
- • Experiments with replacing region proposal network in MaskRCNN with grid based search of Yolo.
- • Panoptic Image Segmentation: Experimental end to end model for instance + semantic segmentation. Uses common feature extraction layer then separate downstream task for scoring, localization and per pixel classification - all trained together.
- • Multi-Video Input: Utilized proprietary data-set to develop methods and types of data that can be distilled from uncalibrated input videos.
End to End Ownership @Faasos @Jugnoo: Owned Product from Story boarding stage to Deployment with Continuous Integration and Delivery, Scaling, and maintainence of Cloud based Services and Architecture.
Faasos Food Services, Mumbai, India.
- • Auth Server @Faasos:”Only server that never crashed”- ex team-lead
- • Created from scratch standalone authentication services for all other applications in the ecosystem.
- • Used JWT for tokens and redis caching for instant delivery
- • Packaged as an middle-ware for all Api calls on diferent backend servers.
- • Inventory Management System @Faasos: "Saved $1mil worth of inventory in 3 years" - ex CTO
- • End to end inventory management system with android application, analytics dashboard
- • keeps track of inventory at every stage of supply chain by reading barcode from android app, updated status tracking for manager on dashboard.
- • Bonus Project: Auto Indentation (warehouse order placement) Timeseries Forecasting @Faasos:
- • Auto Indenting inventory items in supply chain,
- • initial failure with Arima-Shifting to Holtz-Winter gave 87% accuracy.
- • Implemented and delivered as part of Inventory management Tool
- • Database Optimization @Faasos:
- • Reduced 10% deadlocks on Mysql database by partitioning, adding keys, normalization.
- • Shifted non-transactional data to MongoDB.
- • Re-wrote long running transactions as microservices with scheduler caching on dynamoDB and lazy write back to MySQL
- • Devops @Faasos:
- • Created configuration manager services using stateless Amazon Lambda Services and S3 to discover any new service and maintain configurations of all other servers
- • Diverse Tech Stack @Faasos:
- • Developed host of products and maintained a lot more - customer facing application, backend API’s, internal CRM, marketing dashoard, CMS, analytics etc platforms built using ROR, Django, NodeJS, Java,MySql, Redis, MongoDB
- • Self starting Motivation @Faasos:
- • Part of initial 6 member team that updated entire company tech stack from .NeT to MEAN/Django/ROR and scaled to 300k queries/sec transactional database
- • Good Upkeep and maintenance of quality code @Faasos:
- • Refactored huge parts of code base to reduce errors by 25% (measured as delta decrease in number of events logged after commit was merged)
Full Stack Developer
Socomo Tchnologies - Jugnoo, Chandigarh, India
- • God Mode @Jugnoo - Price Surge Model: Developed and deployed single-handedly price surge model algorithm for ride sharing app.
- • Delivers real time update to customer fare depending on ratio of supply to demand, marketing campaigns, seasonal and social trends.
- • Analytic Map shows real-time movement of clients and cabs on single Map Visualization with each trip details, coupons applied etc.
PhD in CS : Computer Vision, Natural Language, & Graphs
SUNY - University at Buffalo
Master of Science in CS: Specializing in Machine Learning
SUNY - University at Buffalo
Bachelor of Engineering in Information Technology
NIT - Institute of Technology