I am Lakshay,
Machine Learning Engineer
and Tech Enthusiast

Hi I am a College graduate with a strong computer science background and experience in Computer Vision and Deep learning. I have 3+ years of Experience in the field and constantly working on upskilling myself.

• Work Experience

Few Beliefs I have in this field

• Data is the king, Data Cleaning is the most important part after gathering data :)
• Taking model accuracy from 90 to 99% is the ultimate task and have to think out of the box to reach there.
• 80-20 rule, 80% of the task takes 20% of the effort, The last 20% requires 80% effort.
• There are many good model architectures out there but sometimes all you need is a small few layers Neural network for your task.
• Many times we don't have to use ML, Image Processing and Software Engineering can avoid the need to use ML.

Machine Learning Engineer
September 2021 - Present

• Aadhaar Masking: Masking Sensitive data on Document: Primary Contributor
• Increased Accuracy of Model from 94% to 99%.
• Reduced P90 TAT from 6 Secs to 2 Secs.
• Reduced Resources Consumption by half.
• Increased Accuracy by using Image Processing and Model training.
• API was scaled to handle 100 RPS and processed 12 Million Docs in 5 days.
• Integrated a flow to add manual flow to reach 100% accuracy.
• Face Liveness: Detect if the Photo taken during KYC is Live or Not?: Secondary Contributor
• One of the biggest selling API of the team with over 1.5M+ requests a month.
• Provide Additional information mandated by RBI like Eye Closed, Blur Detection and Face visibility.
• Increased accuracy from 83% to 99%.
• Reduced TAT from 4 secs to 1.5secs.
• FaceX: Internal API to support other services: Secondary Contributor
• The API aim is to provide fast response to all services that uses Face based processing.
• Reduced dependency on AWS Rekognition API and saved over Rs. 900k.
• Supports use cases for Face Cropping, Face Mask, NSFW, Eye Glasses, Eye Closed etc.
• Super Fast service deployed on Nvidia T4 GPU.
• Used By Liveness, DIV and Tampering service.
• DIV: Identifies the type of the document: Secondary Contributor
• Maintaing and Reviewing Code.
• Updated to new version to reduce cost by 75%.
• The service now handle 1 RPS on a single pod, compared to 0.25 earlier.

• Tampering: Primary Contributor: WIP
• Used ELA To generate dataset for cards.
• Trained those ELA dataset to predict if the card is tampered.

• OCR Services: Maintainer
• Maintaining and Managing tickets for OCR services.

Computer Vision Engineer
January 2020 - September 2021

• Aftom AI
• The system maintains security of staff by monitoring workers through CCTV Cameras.
• Used Nvidia Deepstream to designed a system that was able to take Live multiple input streams process them in real time and shows multiple video outputs.
• Alerts were set and if violated alarms were triggered.
• We were able to process 15 Live streams simultaneously on Nvidia T4 GPU, and advanced algorithms to identify Redzones, used Kafka for communication.
• CV Platform
• Lead and formulated organisation's Computer vision platform where we provided API integrations for services like Face Detection, Mask Detection, Face Matching, ANPR, Helmet Violation, Garbage Detection.
• Anyone can come to website and try our APIs before purchasing them.
• Anti Covid I
• Designed a hardware device that can be plugged into shops and connected with store's CCTV cameras to monitor Covid Protocols.
• Used Jetson Nano and Deepstream to run 4 Live streams that continously sends data to server for dashboards.
• Tracked Face masks, Social distancing and Crowd formations (Hotspots).
• Face based Attendance System
• Online and offline service that uses Live face pics to mark attendance.
• The product was used internally and integrated with CRM to mark attendance of employees.
• It was available as a website and a POC on android was also done.

Get In Touch

If you like to collabrate or wants any advice on any topic, feel free to contact me.