Bhavik

Upadhyay

Developed custom OpenAI Triton kernels across 7 progressive phases: from elementwise ops and reductions to optimized matrix multiply, FFT, convolution, and Flash Attention v2.

• Achieved cuBLAS-parity (within 1%) on tiled matmul at N=4096.
• Benchmarked every kernel against PyTorch/cuBLAS baselines using roofline-model analysis.
• Applied SRAM tiling, L2-reuse group ordering, and memory coalescing.

Constructed core kernel primitives for process scheduling, thread synchronization, and signal handling. Tested end-to-end in QEMU, debugging concurrency hazards across scheduling, VFS, and VM subsystems.

• Enforced process isolation and memory protection via VFS and VM subsystems.
• Enabled kernel-level security guarantees for fork, mmap, and open.
• Devised a split-debugging strategy in QEMU to isolate critical concurrency bugs across scheduling and VFS.

Modular Denoising Diffusion Probabilistic Model trained on CIFAR-10. Implements a full UNet architecture with custom noise scheduling, variance-preserving forward process, and progressive denoising.

• Built a configurable noise scheduler with linear and cosine beta schedules.
• Implemented sinusoidal timestep embeddings and residual attention blocks.
• Achieved class-conditional generation across all 10 CIFAR categories with FID 87.94.

End-to-end robotic manipulation pipeline for the Kinova Gen2 arm, integrating BiRRT motion planning, Task-Space Region grasp sampling, and IKFast inverse kinematics via the AIKIDO planning framework.

• Implemented BiRRT planner with TSR-constrained grasp sampling for reliable object pick-and-place.
• Integrated IKFast analytic IK solver for real-time joint-space trajectory generation.
• Validated full pipeline on physical Kinova Gen2 hardware using ROS and AIKIDO.

Comparative study of three deep RL agents (A2C, DQN, PPO) trained on a custom OpenAI Gym Snake environment, with reward shaping and curriculum scheduling to accelerate convergence.

• Built a custom Gym-compatible environment with configurable grid sizes and rendering.
• Benchmarked A2C, DQN, and PPO; PPO achieved a 60% reduction in training time to target score.
• Reward shaping and curriculum scheduling pushed peak score to 16 points.

End-to-end AI system that identifies coverage gaps in real Python codebases and autonomously writes, validates, and commits passing tests.

• Designed a 6-agent LangGraph pipeline spanning gap analysis, generation, eval, and auto-commit.
• Added multi-stage output filtering before execution as a production-grade LLM reliability guard.
• Sandboxed test runs in Firecracker micro-VMs with full Braintrust LLM call tracing.

Compute-graph pipeline that analyzes PyTorch model topology across five signals to recommend kernel optimizations and hardware placement — grounded in graph structure, no model name or documentation required.

• Each recommendation cites the graph evidence that triggered it and what contradicted it.
• Arithmetic intensity profiled per op type and matched against chip compute ceilings.
• Single trace per model; classification, technique matching, and hardware placement all run weight-free.

Full-stack stock research and paper trading platform: single-page web app, REST API, and native Android client, all deployed on AWS.

• Built real-time stock search and data visualisation pipeline with an Angular SPA backed by a Node/Express REST API.
• Implemented a simulated trading engine with watchlist management and portfolio P&L analytics.
• Delivered a native Android client with feature parity to the web app, deployed across AWS EC2 and S3.