NS
Nayant Srivastava
02ENGINEERING / 02 · REPOSITORY DIRECTORY

Systems & Research Portfolio

Engineered systems evaluated against quantitative benchmarks, hardware constraints, and failure modes.

Filter Tier:
TIER S · FLAGSHIP2025 — Present
Research Prototype / Benchmark Ready

ADCTM: Autonomous Data Centre Thermal Management via Closed-Loop RL

Coupling Fourier Neural Operator heat-flux surrogates with safe reinforcement learning to eliminate thermal hotspots and reduce data-centre cooling energy by 28.4%.

Cooling Energy Reduction28.4%
Technologies
Safe Soft Actor-Critic (SAC)Fourier Neural Operators (FNO)PyTorchOpenAI GymnasiumOpenFOAM CFDNavier-Stokes Convection PDEsReduced-Order ModelsNumPyPython C-ExtensionsZeroMQ Telemetry BusSIMD Tensor KernelsTracy Profiler
Read Full Case Study
TIER S · FLAGSHIP2024 — 2025
Validated Research Prototype

Autonomous Self-Aware Spacecraft: Fault-Tolerant Neural Subsystem

A sub-millisecond edge neural architecture enabling deep-space probes to detect structural anomalies, diagnose cascading component failures, and adapt control policies under strict radiation-hardened compute constraints.

Inference Latency4.2 ms
Technologies
Temporal Graph Neural NetworksExtreme Value Theory (EVT)Unsupervised Anomaly DetectionPyTorchEmbedded C99 / RustARM Cortex-M4 (Flight Emulation)Zero-Allocation HeapSPI / CAN Bus DriverSliding-Window Peaks-Over-ThresholdKalman Filter EnsemblesSpectral Eigen-Analysis
Read Full Case Study
TIER A · MAJOR2025 — Present
Active Development

MINDLITE: High-Throughput Machine Learning Experimentation & Quantized Runtime

A zero-dependency C++20 inference engine and reproducible experimentation harness designed for rapid iteration on resource-constrained hardware.

Inference Speedup3.4x
Technologies
C++20SIMD Intrinsics (AVX2/NEON)OpenMPCMakePyTorchONNX ExportPost-Training QuantizationAblation Harness
Source Code
TIER A · MAJOR2025 — Present
Development

Omnix Engine: High-Performance 3D Simulation & Rendering Pipeline

A custom ground-up C++20 graphics engine using modern Vulkan 1.3 to achieve sub-2ms frame times for large-scale dynamic particle and physics simulations.

Frame Time< 1.8 ms
Technologies
Vulkan 1.3GLSL / SPIR-VBindless DescriptorsVMAC++20Data-Oriented ECSLinear Memory AllocatorsTracy Profiler
Source Code
TIER A · MAJOR2025 — Present
Research Prototype

ASTRA: Adaptive Spatial-Temporal Telemetry Reasoning Architecture

Graph-structured temporal reasoning for complex distributed multi-sensor telemetry networks under high-packet-loss environments.

Packet Recovery99.4%
Technologies
Dynamic Graph AttentionMatrix ImputationKalman FiltersZeroMQProtobufC++20Linux Network Stack
Source Code