NS
Nayant Srivastava
Open to AI/ML & Systems Roles

NAYANT SRIVASTAVA

AI / ML Engineer

Reinforcement Learning + Intelligent Systems

Nayant Srivastava — AI / ML Engineer

I build intelligent systems and explore the engineering problems behind them.

Foundations in Systems Engineering, DSA, Distributed Architecture & Applied Mathematics
View ProjectsResume / CVGitHub
Research & Engineering Focus:
Autonomous Telemetry Anomaly ReasoningTemporal Graph NetworksOffline Reinforcement LearningQuantized Edge Execution
01Status // Real-Time

Current Focus

A persistent log of active engineering work, ongoing research questions, and technical explorations.

Building
Autonomous telemetry anomaly reasoning engine using Temporal Graph Networks with quantized edge execution.
Researching
Sample-efficient offline reinforcement learning under non-stationary dynamics and distribution shifts.
Learning
Distributed consensus algorithms (Raft internals) and Linux eBPF telemetry hooks for low-overhead model observability.
Contributing
Gymnasium benchmark environments and memory-profiling tooling for PyTorch model inference.
Open to
AI/ML Engineering, Intelligent Systems Architecture, and High-Performance ML Systems opportunities.
02Work // Evidence

Selected Projects

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

TIER S · FLAGSHIP6 Months · Fall 2025 — Present
Active Research / v1.2 Prototype

Self-Aware Spacecraft: Telemetry Anomaly Detection & System Reasoning

Autonomous anomaly detection and root-cause reasoning across 1,200+ telemetry streams under extreme edge latency constraints.

False Alarm Reduction64%
Technologies
PyTorch 2.4Temporal Graph Networks (TGN)Quantized INT8 PyTorchContrastive Representation LearningC++20 EngineONNX Runtime EdgeZeroMQ Telemetry BusLinux eBPF TracingSIMD IntrinsicsNASA SMAP / MSL DatasetsHigh-Rate Synthetic Fault InjectorPrometheusGrafana Telemetry Sink
Read Full Case Study
TIER A · MAJOR4 Months · Summer 2025
Research Preview / In Development

Animath: Differentiable Physics & Neural Operator Engine

Solving high-dimensional non-linear PDE boundary value problems 400x faster than traditional numerical solvers using Fourier Neural Operators.

Solver Speedup420x
Technologies
PyTorchFourier Neural OperatorsJAXDifferentiable PhysicsSymplectic IntegratorsCUDA C++HDF5Eigen3
[Case study in documentation]Source Code
TIER A · MAJOR3 Months · Spring 2025
v0.9 Prototype

Omnix: High-Throughput Distributed Feature Store for Real-Time Inference

Sub-millisecond distributed feature store and streaming aggregation engine engineered for high-concurrency real-time ML inference.

p99 Read Latency0.84 ms
Technologies
RustTokio AsyncRaft ConsensusgRPC / ProtobufRocksDBMemory-Mapped FilesLinux epollPrometheus
[Case study in documentation]Source Code
03Methodology // Core

Engineering Philosophy

Distinguishing disciplined systems engineering from speculative experimentation.

// 01From theory to systems

Build

Turn mathematical formulations and model architectures into resilient, production-grade systems rather than leaving them as isolated notebook prototypes.

// 02Empirical over intuitive

Measure

Don't rely on intuition when an empirical experiment, profiler trace, loss curve, or ablation study can answer the question with certainty.

// 03Search for failure modes

Break

Actively stress-test boundaries, perturb input distributions, simulate adversarial telemetry, and probe edge cases to uncover latent failure modes early.

// 04Full-stack mechanical sympathy

Understand

Investigate why systems succeed or degrade — from algorithmic gradient flows and loss formulations down to CPU/GPU memory caches and network bandwidth.

// 05Evidence-driven iteration

Improve

Iterate systematically based on verified quantitative benchmark data and disciplined root-cause diagnoses, never speculative premature optimization.

04Trajectory // History

Experience & Journey

Formal engineering roles, technical research positions, systems milestones, and competitive achievements.

2025 — Present·Research

Lead ML Systems Architect

@ Self-Aware Spacecraft Project

Designing real-time anomaly detection and causal reasoning engines for high-dimensional satellite telemetry streams under sub-20ms latency limits.

PyTorchTemporal GNNC++20ONNX RuntimeeBPF
2024 — 2025·Role

AI / ML Engineering Fellow

@ Autonomous Agents Lab

Engineered reinforcement learning training pipelines, reward shaping frameworks, and vectorized environment rollouts for complex continuous-control tasks.

Reinforcement LearningGymnasiumRay TuneDistributed Training
2023 — 2024·Research

Systems & Algorithms Researcher

@ High-Performance Computing Group

Investigated cache-oblivious algorithms, high-throughput feature caching layers, and parallel graph partitioning algorithms for large-scale data structures.

DSAC++System DesignLinux InternalsConcurrency
2023·Hackathon

First Place Winner

@ National AI Systems Hackathon

Constructed an edge computer-vision pipeline running localized inference on constrained hardware with real-time audio-visual feedback loops.

Edge MLTensorRTEmbedded LinuxReal-time Systems
05Community // Impact

Contributions

Open-source optimizations, technical writing, benchmark reproductions, and developer tooling.

Open SourceMerged

Graph Tensor Batching Optimization

Contributed sparse graph tensor collation and memory-efficient batching routines to open-source ML acceleration utilities, reducing allocation overhead by 28%.

View PR & Code
ResearchReproduced

Offline RL Benchmark Reproductions

Validated and reproduced Conservative Q-Learning (CQL) and Decision Transformer baselines on continuous control benchmarks under distribution drift.

Technical Report
Technical WritingTechnical Note

Understanding Memory Latency in GNN Inference

Published an engineering post analyzing cache misses, sparse adjacency matrix traversals, and quantization strategies for on-device edge ML.

Read Article
Developer ToolsTool

vram-trace: Lightweight GPU Memory Profiler

Engineered a zero-overhead CLI profiler that tracks CUDA tensor allocations, fragmentation events, and memory leaks during PyTorch RL training runs.

GitHub Repository
06Identity // Depth

About Me

Four layers of professional identity, engineering lineage, cognitive model, and technical trajectory.

01.

Who I am

/Professional Identity

I am an AI/ML Engineer focused on Reinforcement Learning and intelligent systems. I specialize in building computational architectures that can reason, adapt to non-stationary environments, and execute reliably under strict physical and latency constraints.

02.

How I got here

/Background & Path

My foundation was built in competitive programming, algorithms, and systems engineering. When I delved into machine learning, I recognized that deploying intelligent models into production is primarily a systems problem: managing memory bandwidth, ensuring telemetry integrity, and handling real-world distribution shifts.

03.

How I think

/Engineering Philosophy

I prioritize depth first and breadth second. Rather than treating ML as a black box of off-the-shelf APIs, I seek mechanistic explanations down to the loss gradient, matrix formulation, and CPU/GPU memory cache. Every engineering claim must be backed by reproducible experiments and metric baselines.

04.

Where I'm going

/Future Trajectory

I am focused on designing autonomous, self-healing systems that close the loop between perception, causal reasoning, and real-time control — particularly for aerospace, robotics, and distributed infrastructure where failure is not an option.

07Channel // Contact

Get In Touch

Direct contact channels for engineering inquiries, research collaborations, or technical opportunities.

Direct Coordinates
nayant@example.com
github.com/nayantsrivastava
linkedin.com/in/nayantsrivastava
Typical response latency: < 24h for technical & recruitment inquiries.
Message Transmission[v0.1 Mock Handler]