Backend Engineer. Distributed Systems. Problem Solver.
Measured in production, every metric below came from a live system at Siemens, processing 1M+ requests a day.
Four years in production backend systems, and a research lineage that started before that — now pointed at AI agents. Same instinct: understand the system, then build it.
From building recommendation systems at NVIDIA to optimizing latency on backend services handling 1M+ requests a day — every stop added a layer.
Own backend infrastructure for distributed services on Java & Spring Boot — distributed locking via Redisson RLock, denormalized policy data access, SNS/SQS event-driven sync with DLQs, and zero-downtime schema migrations on MySQL using Percona pt-osc.
- Implemented distributed locks via Redisson RLock, 1M+ req/day, eliminating 5–6% concurrent overprovisioning under burst traffic.
- Redesigned policy data access using a denormalized store, removing a 10K+ ID lookup bottleneck and cutting latency by 60%.
- Architected SNS/SQS-based event-driven sync for policy store consistency, with DLQs and configurable retries for fault tolerance.
- Led zero-downtime schema alteration on a 30M+ record table using Percona pt-osc, avoiding 3 hours of production downtime
- Reduced POST API latency by 20% via multi-threaded validation pipeline using the observer pattern for large payloads.
Developed interactive data grid components with filtering and pagination in Angular. Implemented chart rendering and notification popup features in dialog boxes, resolving 6 UI bugs across the application.
- Developed an interactive data grid components, with filtering and pagination in Angular Framework.
- Implemented chart rendering and notification popup features in Angular dialog boxes; resolved 6 UI bugs across the application.
- Added JUnit tests and resolved SonarQube flagged bugs, code smells, improving overall code quality by 15%.
Trained DLRM architecture on the Criteo 1TB dataset, achieving 0.639 accuracy over 10K iterations using Merlin and NVTabular. Implemented feature engineering and embedding layers on raw clickstream data.
- Trained DLRM architecture on Criteo 1TB dataset achieving 0.639 accuracy over 10K iterations
- Implemented feature engineering techniques, embedding layers, on raw clickstream data
- Benchmarked 3 popular recommendation system architectures against DLRM on GPU, published comparative training statistics as blogs.
Worked in the field of Natural Language Processing at Centre of Excellence in Safety Engineering and Analytics, IIT-KGP.
- Worked in the field of Natural Language Processing at Centre of Excellence in Safety Engineering and Analytics, IIT-KGP.
- Analyzed Construction Site Catastrophe reports by applying chunking, TF-IDF vectorization, word embeddings to classify causes of accidents and chunk out fatal objects.
- Proposed a Hybrid Neural Network Architecture with Attention Layers with an F1 score of 0.88.
Worked across computer vision, RNNs, and data science — predicting cloud failure on Google Cluster Traces data, and building a compact face-mask detection model optimised for low latency.
- Working in the field of Computer Vision, CNNs, Data Science at Corporate Venturing and Innovation Group, Tata Communications.
- Worked on building a compact Face-Mask Detection model with low latency.
- The proposed model outperforms VGG-16, MobilenetV2 architectures.
Foundation in algorithms, data structures, and systems thinking — the launchpad for everything that followed. Spent the latter half of college alternating between ML research internships and building open-source projects.
- Cracked multiple off-campus internships at Big Techs that shaped how I think about systems at scale from scratch from the very begining.
- Built two open-source projects - Sign Language to Speech, and Blind Assistance System.
- Published research papers along coursework.
Two open-source projects with real community traction, three published research papers — built before I had a job, still cited and starred today.
Real-time object detection for visually impaired users using TensorFlow Object Detection API and SSD architecture on a live webcam feed, achieving 98% accuracy. Estimated object proximity using bounding-box area ratio and delivered real-time audio alerts — published in Springer Journal.
CNN-based sign language classifier using TensorFlow translating static ASL gestures to speech output. Published in IJNGC.
Proposed the CCNet model for construction site accident classification — a deep learning architecture that classifies accident types and root causes from visual site data.
Trained DLRM architecture on the Criteo 1TB dataset, achieving 0.639 accuracy over 10K iterations using Merlin and NVTabular. Implemented feature engineering and embedding layers on raw clickstream data.
I'm a backend engineer who builds distributed systems himself — and is now extending that into AI agents. If you're working on something that needs to handle real load, or a team exploring agentic AI, I'd love to hear about it.
Open to conversations with engineers and teams doing interesting work.
Fill this out and hit Send Message to open it in your mail app.