Satyam Kesharwani

Ministry of Defence–sponsored project, National Institute of Technology Kurukshetra · Kurukshetra, India

• Sole student developer on an MoD-sponsored security project; built a security analysis suite of 46 interconnected modules to detect and mitigate threat vectors, following secure-coding practices.
• Analyzed high-volume firewall, antivirus, DNS and Windows event logs to model user behavior and flag threats; followed SDLC practices through design, implementation and testing.

National Institute of Technology Kurukshetra · Kurukshetra, India

• Built an offline battery-diagnostics system on an ESP32 whose Residual-Physics Neural Network reaches R² = 99.70% (XGBoost: 98.46%); second author, IEEE NE-IECCE 2026.
• Compiled the PyTorch model to INT8 C++ with the Edge Impulse EON Compiler for fully offline inference on the ESP32.

Read the write-up

National Institute of Technology Kurukshetra · Kurukshetra, India

• Fine-tuned a pretrained Vision Transformer (frozen ViT-B/16 backbone, new classification head) on images built from sandbox behavior reports of 4,280 Android apps: 99.78% accuracy for ransomware detection; lead author, ICDAM 2025 (Springer).
• Built the Python pipeline from CuckooDroid reports to features and images; compared CNN and Random Forest baselines.

Read the write-up

AtomicKV: Distributed Key-Value Database

C++17, Linux epoll, B-Tree, Gossip, AWS EC2

• Built a production-grade distributed database from scratch on a single-threaded Linux epoll event loop.
• Tiered storage: an LRU cache plus a Bloom filter for instant misses, backed by a custom on-disk B-Tree.
• Consistent hash ring with virtual nodes, masterless gossip failure detection, async replication with read repair and anti-entropy, and Lamport clocks for last-write-wins conflict resolution.
• Load-tested with 200 concurrent clients (80K requests): 10,000+ req/s at ~16 ms average latency (p99 ~40 ms); deployed on AWS EC2.

Source codeWrite-up, part 1Write-up, part 2

Custom Memory Allocators & Order Matching Engine

C++17

• Wrote Linear, Stack, Pool and Free-List allocators from scratch and used the pool in a price-time-priority limit order book, so the matching path never calls new/malloc.
• 1M alloc/free operations (median of 7 runs, -O2): pool ~2.5× and linear ~15× faster than glibc new/delete.

Source codeWrite-up

Shortify: URL Shortener with Multi-Layer Caching

Node.js, PostgreSQL, Redis, Docker, AWS

• In-process LRU and Redis caches over PostgreSQL, a Bloom filter for unknown codes and a Redis lock against cache stampedes: 7.12 ms average redirect latency under k6 (50 virtual users, 100% success).
• Base62 Snowflake IDs, a Redis-backed token-bucket rate limiter, Docker deploys to AWS EC2 through a GitHub Actions CI/CD pipeline.

Source codeWrite-up