Hemang Shah Email & Phone Number

Bengaluru, Ka, In

Building powerful AI tools for consumer insights teams, market researchers & brand design professionals.

Bengaluru Area, India

Streamoid is a Fashion AI company (https://www.streamoid.com), provides products for fashion retailers.I've helped build the following:AI Fashion Stylist Chatbot: using LLMs like ChatGPT, incorporates Streamoid's internal fashion knowledge, integrates deeply into client's inventory and recommends products based on current conversation context. (https://www.streamoid.com/vortex)Catalogix: A custom microservices orchestration pipeline to manage various ETL and automation workflows. Cloud-native, auto-scalable, configurable, easy to add new workflows, and having high visibility.(https://www.streamoid.com/catalogix)Outfitter: A fashion recommendation system that uses Streamoid's internal fashion knowledge to find products that can be paired with the hero product, to create a complete outfit. (Given a shirt, to find a pair of trousers, shoes, jacket, etc that would make a coherent outfit)Powers widgets like 'Complete the look' or 'Shop the look' on a Product Display Page.In addition to this, drove various internal tools and frameworks for structuring product data, for fashion team to provide intelligence, etc.(https://www.streamoid.com/insearch)Matcher: Another fashion recommendation system that finds multiple products in complementary categories that would pair well with the hero product.(https://www.streamoid.com/insearch)Various misc: Virtual try-ons, segmentation, background generation using Stable Diffusion, a bunch of internal systems, etc.

Research and development in the domains of computer vision, machine learning and algorithm development.Project 1: To determine clothing attributes such as type of clothing worn (top or dress, pant or skirt), lengths of clothing (length of sleeves, length of skirt/pant), etc, given a full body still image of a person taken in natural pose and natural environment.Detection of head-shoulder: Machine learning algorithm using HoG descriptors.Segmentation of human body: Using generative modelling of background and foreground using mixture of gaussians.Segmentation of clothing parts: Roughly segmenting which regions are torso, mid, legs, etc, by employing several classifiers based on fisher's discriminant analysis and support vector machines, as well as fitting to linear models using linear regression.Classification of type of clothing, length, etc: Using classifiers and linear regression.Developed libraries to work with sets of pixels and sets of neighbouring regions; to work with sets of colors and problem-specific methods of determing color difference, as well as algorithms for merging, splitting and modelling color groups; for extracting geometrical features; for pyramid-like segmentation using graph-cuts; for bounded hierarchical clustering, etc.Lastly, a system of modelling a set of colors to represent a segmented clothing item, and a fast image retrieval system using nearest-neighbors to fetch items similar to segmented clothing was built.Software was developed in C++ and was designed for online processing (as opposed to offline).

Bengaluru Area, India

Worked on sZensEYE, a navigational aid for the visually impaired. sZensEYE is a low cost monocular vision device that detects obstacles and communicates their direction and proximity to the user.

Bengaluru Area, India

Vidgyor enables broadcasters to monetize their internet TV streams by playing targeted video advertisements in place of broadcast TV commercials.

Bengaluru Area, India

Worked on iorad, a tool for creating screenshot based tutorials. Developed novel algorithm for understanding ui context of a click on a screenshot.

ViBE is the leader in change detection / background subtraction algorithms.Change detection or background subtraction refers to detection of foreground pixels in a video sequence.It is a fundamental requirement in many many video processing applications. Some verticals are video surveillance and video content analysis, user gesture tracking, people/vehicle/object tracking and counting, room occupancy monitoring / parking occupancy detection, activity localization, low bandwidth video streaming, stolen / abandoned object detection, privacy in video, etc.Please do feel free to connect with me to know more about ViBE or about how change detection algorithms in general can speed up your application.

Part of 3 member core team. Lead several technical projects, contributed in business ideation, strategy and fund raising, hired and managed engineers, developed several technology IPs & multiple patentable ideas.Excerpts:Mixed Reality: Facial expressions from a webcam stream are transferred onto a portrait image.This project involved: face tracking using active appearance models, accurate face contour determination, transfer of facial expressions, transfer of facial color, restoration of hidden body parts such as mouth, neck, blending.Projection Correction: Using stereo vision to:Correct color errors introduced by color of screen on which a projection is made and errors by ambient lighting.Transform shape of projection into a rectangle while being projected at an angle to the screen.Interactive human body segmentation.OCR pre-processing - process noisy color images for standard OCR recognitionDetection, tracking and replacement of proprietary Augmented Reality marker.Color Code:Improve algorithms for fast detection and decoding of a 2D color code (square grid of colors)Improve quantization algorithms for correct classification of color code under strong ambient light, printing errors and colored paper.Misc computer vision:Recognition of hand gestures, video stabilization, ffmpeg filter development, object recognition, face detection & tracking, face recognition, face authentication & fingerprint recognition.Recommender System:Developed a campaign management tool for advertisers to serve relevant contextual recommendations to users based on their purchase histories.Networking:Developed peer-to-peer audio & video communication for a 250MHz Arm9 device using RTP over SIP (using pjsip). NAT traversal was done using STUN, TURN protocols, as well as udp hole punching techniques.Misc:Enhanced an open source diff algorithm for determining inline differences within a block, as well as correct alignment of lines in a block for comparison.

Co-founded Blink Media Pvt. Ltd. - a media / consumer electronics / analytics startup – which created the ‘intelligent shopping cart’, a sophisticated touch-screen device placed on the handle of a shopping cart in a supermarket, that engages shoppers and brands by providing highly contextual targeted content (like ads, recommendations, tips, recipes), based on cart’s location in the store and the shopper’s purchase history and current basket.Software developed includes that for the device application, server software, network architecture, shopper analytics solution, recommendation engine, web application for campaign management, etc.Product demo:www.youtube.com/watch?v=5psujp_xduUNews Article: economictimes.indiatimes.com/News/News-By-Industry/Services/Retailing/Best-offers-on-shopping-cart-at-Reliance-Retail/articleshow/3395131.cms

Software design & development in C++.Distributed systems, OLAP, high performance in-memory data analysis, in-memory data organization.