Piush Paul — Systems & Software Engineer

C++20 / CMAKE DPDK FAST-PATH HFT MATCHING

High-frequency cryptocurrency order book & kernel-bypass execution engine

Architecting a low-latency cryptocurrency HFT engine with a zero-allocation feed handler, L2 order book, and execution engine targeting sub-microsecond latency.
Optimized critical path latency and inter-thread communication by implementing bounded lock-free MPSC ring buffers, core pinning (thread affinity), cacheline alignment, and branch prediction hints.
Engineered a custom WebSocket networking layer via DPDK kernel bypass, routing packets to userspace to bypass OS network stack overhead.
Zero-allocation TCP/IP state machine and direct NIC polling to eliminate OS context switches and interrupt overhead.
Zero-copy payload extraction for market data feeds, slashing tail latency during burst conditions.

MODERN C++20 LOCK-FREE ATOMICS 1.08B OPS/SEC

Two-tier thread-cached memory allocator with 14 ns p50 latency

Architected a two-tier thread-cached memory pool (thread-local L1 cache + central slab), scaling throughput to 1.08 Billion ops/sec across 12 concurrent threads (808× speedup over global spinlocks) with a 14 ns p50 allocation latency.
Engineered an ABA-free lock-free allocator utilizing a 64-bit packed TaggedIndex atomic CAS (lock cmpxchgq) and dense atomic index array, completely eliminating lock contention with zero data races verified under TSan and ASan.
Optimized memory subsystem with strict 64-byte cacheline alignment (alignas(64)), pre-faulted 2 MB Huge Pages (MAP_HUGETLB), and software prefetching, eliminating false sharing, runtime TLB thrashing, and OS kernel page faults.

AGENTIC AI GOOGLE ADK GEMMA & GEMINI

Hybrid edge/cloud agricultural intelligence • 1st Runner Up Google AI Hackathon

Architected a hybrid AI mobile app for the ~70% of rural farmers lacking reliable connectivity, integrating on-device Gemma for offline multilingual voice Q&A with Gemini on Vertex AI for multimodal crop disease diagnosis.
Engineered Python telemetry pipelines to analyze Sentinel-2 satellite data across 6 spectral indices (NDVI, NDWI, EVI), delivering crop yield predictions to guide harvest planning.
Awarded 1st Runner Up at the Google Agentic AI Hackathon 2025, which held the Guinness World Record for the most participants in an agentic AI hackathon.

ISAI 2025 PAPER EFFICIENTNETB2 96.12% PRECISION

Advanced Deep Learning Techniques for Early Detection of Oral Cancer: A Multimodal Approach

Published in Second International Symposium on Artificial Intelligence (ISAI 2025) by Piush Paul, Umang Singh, Sarah Khan, Atharva Jog, Ujwala Bharambe.
Developed and optimized deep learning models using advanced data processing techniques on the Oral Cancer Image dataset, enabling robust performance comparison across CNN, EfficientNetB2, and Vision Transformer architectures through rigorous metric evaluation.
EfficientNetB2 achieved state-of-the-art oral cancer classification performance with 96.12% precision and 94.38% accuracy, significantly outperforming baseline CNN (76.61%) and ViT (83.69%), demonstrating strong clinical potential.

GEE & SENTINEL-2 CNN-LSTM MSE: 0.0069

Spatiotemporal vegetation forecasting & pollination hotspot modeling

Integrated and preprocessed a 4-year, multi-source dataset of satellite imagery and climate data (Sentinel-2, MODIS, NASA SRTM), applying temporal interpolation to correct for cloud cover.
Trained a CNN-LSTM spatiotemporal model to forecast vegetation changes and developed a custom algorithm to classify blooming stages with high accuracy (MSE: 0.0069).
Implemented K-Means clustering to identify optimal beehive "hotspots" based on key environmental factors, delivering weekly visualized recommendations for apiculture yield.

NEXT.JS SSR FLUTTER FIREBASE REALTIME

HelpAlert disaster response & Vendo municipal street vendor monitoring

Built HelpAlert: Disaster relief management platform combining Next.js SSR portal for civic authorities with Flutter mobile client for citizens; Google Maps safe zones and Firebase Authentication.
Engineered Vendo: Municipal street vendor governance platform enabling authorities to manage weekly bazaars, track vendor permits, and coordinate vending zones via dynamic SSR and Axios.
Secured Winner status at Smart India Hackathon 2022 (Government of India) & Codessiance Hackathon.

FEATURED PROJECTS