Akshay Kulshreshtha

Akshay Kulshreshtha

better than your llm inference engineer

// didnt have time to build the UI, cause i was busy optimizing the kernels

Experience

AI Automation Intern, HiddenMindSolutions

Sep 2025 – Nov 2025

  • Built and deployed an end-to-end customer support automation system using n8n, integrating Email/WhatsApp with Zendesk/Freshdesk APIs, handling 100+ tickets/day and reducing manual effort by 80%.
  • Developed AI-powered voice calling agent using ElevenLabs to automate receptionist tasks for a California-based golf club, enabling real-time tee time booking with calendar integration and reducing front-desk workload by ~70%.
  • Designed and optimized prompt-driven AI agents for automation and conversational workflows, improving response accuracy, contextual understanding, and system reliability.
View experience letter

Starred Projects

Memory-Efficient LLM Inference via Quantized KV Cache Compression

  • Implemented polar quantization + 1-bit QJL residual correction for LLM KV cache; achieved 3.12× compression (68% memory saving) with 0.9967 cosine similarity.
  • Engineered fused Triton kernel with 2D block mapping (BLOCK_M=16) to trigger Tensor Cores, delivering 2.26× speedup over PyTorch baseline.
  • Technologies: PyTorch, Triton, Kubernetes, Docker, Pydantic, Hugging Face Transformers, FastAPI
KV cache memory (relative) Baseline 100% Compressed 32% Speed vs PyTorch baseline PyTorch 1.00× Fused Triton 2.26×

3.12× compression while keeping 0.9967 cosine similarity with the original KV cache.

Fault-Tolerant Distributed Training System for LLM Datasets

  • Built a fault-tolerant multi-GPU training pipeline using PyTorch DDP, reducing training time by ~50% with efficient checkpoint-based recovery.
  • Eliminated distributed training bottlenecks through communication and memory optimizations, achieving 92% BERT-base accuracy while improving scalability across GPUs.
  • Technologies: PyTorch, Distributed Data Parallel (DDP), CUDA, Prometheus, Python.
Time to reach target accuracy target accuracy ~50% less training time Training time Progress to target Fault-tolerant Baseline

Illustrative sketch of the ~50% training-time reduction, not measured data.

see more on github

Papers & Preprints

Momentum-Aware Reinforcement Learning for Free-Floating Manipulation in Microgravity

  • Developed a PPO-based curriculum learning framework for 13-DoF free-floating space robot assembly, structuring long-horizon manipulation into a 5-stage pipeline under momentum constraints.
  • Introduced a momentum-aware reward function reducing peak angular momentum by 42.6% and achieving a 78.6% success rate in zero-damping PyBullet simulations.

Self-Supervised Neural Architecture Search in Low-Data Regime for Financial Time Series

  • Designed a two-stage SSL+NAS framework combining contrastive representation learning (InfoNCE) with DARTS to automatically discover causal convolutional architectures for financial time-series prediction in low-data regimes.
  • Achieved 64.49% test accuracy and positive risk-adjusted performance (Sharpe 0.91) across multi-stock datasets using only 1,000 training windows, outperforming supervised LSTM and Conv1D baselines.

Skills

Programming Languages

  • Python
  • Triton

Core Machine Learning & AI Concepts

  • Machine Learning
  • Deep Learning
  • Natural Language Processing
  • Transformers
  • Large Language Models

Framework & Libraries

  • PyTorch
  • PyTorch DDP
  • PyTorch FSDP
  • DeepSpeed

Tools & Technologies

  • Kubernetes
  • Docker
  • Git
  • GitHub
  • Prometheus
  • Grafana
  • Hugging Face Transformers
  • n8n
  • Pydantic
  • Redis
  • Celery
  • GitHub Actions
  • Terraform

Education

Vellore Institute Of Technology, Bachelor of Technology in Computer Science with Specialization in Artificial Intelligence & Machine Learning.

2023 – Present

  • GPA: 7.60/10

Certifications

IBM: Generative AI using IBM watsonx

Earned 20+ skill badges from Google Cloud

MATLAB Onramp from MathWorks

Achievements & Leaderships

  • Vice President, Hindi Association, VIT-AP: Led 10+ cross-functional teams (20+ members each) to organize events for 200+ participants; partnered with 5+ clubs to boost diversity and turnout.
  • Team Member, CodeChef VIT-AP: Organized intercollegiate coding contests (15+ colleges), provided technical support, maintaining a 95% issue resolution rate and resulting in an increase in repeat participation by over 20%.