Headline

ML Systems @ EatClub Brands (Box8) | NIT Allahabad

About

Over the past ~4 years, I’ve designed and deployed end-to-end ML systems across computer vision and time series use-cases — from identifying business problems to shipping production-grade services. My work spans: • Real-time computer vision pipelines • On-device and edge ML deployments (Android + desktop) • Cloud-native ML systems on AWS & GCP • Designing data-intensive systems with clear tradeoff decisions I care deeply about: • Systems thinking over model obsession • Benchmarking and measurable performance • Building things that actually work outside notebooks Currently building: Focusguard — a real-time, on-device posture & fatigue estimation system. Technical stack: Python, C++, Docker, Linux, AWS, GCP Comfortable with Postgres, Kubernetes, Kibana, and system-level debugging. I’m particularly interested in problems involving real-time perception systems, edge intelligence, and applied ML research.

Location

Bengaluru

AI stack

claude code

gemini

hugging face

ollama

opencode

Jina AI

Work experience

Applied ML Engineer at EATCLUB Brands (Formerly BOX8), Jun 2022 - Present · 4 yrs 3 mos, Bengaluru, ◦ Computer Vision (Edge): Developed a low-cost attendance system involving anomaly detection using computer vision for 250+ outlets across India, saving $20k in operational cost. ◦ Computer Vision (Cloud): Developed, deployed and monitored 8 CNN based Object Detection, Image Classification and Instance Segmentation pipelines using Apache Airflow and Kibana, which process 40k images per day and flag critical business issues with 95% accuracy, bringing down manual audit time to 1/4th of original. ◦ Time series forecasting: Used Prophet to predict the usage of critical resources in the Ops and Cart microservices, leading to a 5% cost reduction and lesser downtime durations. Performed EDA, data preprocessing and cross validation to bring down the Peak MAPE below 10%. ◦ Clustering: Applied DBSCAN on address embeddings encoded by RoBERTa Tokenizer for the creation of PoI Polygons, leading to a reduction in our last mile delivery time and associated fuel cost. ◦ ML Lifecycle Experience: Involved end-to-end from data wrangling to labelling, model training, deployment, metric evaluation, monitoring and iterative enhancement. ◦ Projects: Predictive Application Scaling • PoI Mining via Address Embeddings • Human grooming detection • Seven segment weight detection • Kitchen order ticket detection • Burnt pizza detection • Ingredient detection • Device detection • Number plate recognition • Camera Inspection Models I've worked with: - Time series forecasting: - Prophet, LSTM - Unsupervised Learning - DBSCAN - Object Detection, Image Classification, Instance Segmentation: - YOLO (You Only Look Once) - v8 - SSDMobNetV2 - Single Shot MultiBox Detector - Faster R-CNN (Region-based Convolutional Neural Networks) - OCR: - PaddleOCR, EasyOCR, TesseractOCR

Data Structures and Algorithms Mentor at The 10x Academy, Sep 2021 - Jun 2022 · 10 mos, Remote, • Interviewed 300+ students in 1:1 approach correction mock sessions involving improvisation of conceptual understanding and discussion of good problem solving practices. • Conducted multiple group discussions related to problem solving involving various data structures and algorithms. Also helped students during practice sessions to move towards the most optimal solutions.

SWE Intern at Glabbr, Mar 2021 - May 2021 · 3 mos, Pune, Maharashtra, India, • Worked on creating, testing and deploying RESTful APIs for microservices using SRP. • Was involved in the design and POC of cypher queries on neo4j as well as their subsequent integration into the microservice. Additional Tools Used: neo4j, Dgraph

Education

Bachelor of Technology, Motilal Nehru National Institute of Technology

Public links

github: https://github.com/prateek-mehra

leetcode: https://leetcode.com/prateek_mehra