Ashwattha Phatak

Conflict-aware locking model for shared multi-agent memory using semantic similarity instead of key equality.

Technical focus: Distributed locking extensions beyond key-space conflicts through semantic similarity and commit-time validation.

• Benchmarked semantic conflict detection quality against key-level locking baselines.
• Measured throughput behavior under varying contention levels and lock scopes.

Stack: Python, FAISS, Vector Search, Distributed Coordination

Desktop app that sanitizes identity images with local ML defenses to reduce deepfake manipulation risk.

Technical focus: End-to-end integration across desktop runtime, frontend UX, and local ML inference in a privacy-first workflow.

• Implements 3 user-selectable protection levels (Low, Medium, High).
• Supports desktop packaging for macOS, Windows, and Linux.

Stack: Next.js, Tauri, Rust, Python, PyTorch, OpenCV

ROS2 + Gazebo autonomy stack with multi-agent exploration and reinforcement-learning policy evaluation.

Technical focus: Autonomy pipeline integration across localization, planning, and control under noisy sensing and constrained compute.

• Compared exploration coverage behavior against baseline planners in simulation.
• Measured localization stability and drift patterns across filter configurations.

Stack: Python, ROS2, Gazebo, Nav2

Linux memory-access profiling pipeline that combines eBPF sampling with automated visualization for program behavior analysis.

Technical focus: Linux observability and low-level profiling, translating kernel-level signals into actionable diagnostics.

• Profiles page activity with a 25 ms sampling/export window for timeline analysis.
• Generates top-page visualizations (default top 512 pages) for workload comparison.

Stack: C++, eBPF, libbpf, Python, Linux

WAFL-inspired distributed filesystem with copy-on-write snapshots, journaling, and replicated metadata.

Technical focus: Correctness-critical filesystem behavior under concurrent writes, snapshotting, and crash recovery.

• Benchmarked snapshot overhead under mixed read/write workloads.
• Validated recovery behavior through repeated crash and restart fault-injection tests.

Stack: C++, Linux, Multithreading, Sockets

Range-image compression/reconstruction experiments for 3D detection latency-accuracy tradeoff analysis.

Technical focus: Perception experiment design focused on turning model-level results into deployment-oriented guidance.

• Characterized latency and AP/recall trade-offs across multiple input resolutions.
• Measured runtime variance and detector stability under compressed input conditions.

Stack: OpenPCDet, PyTorch, KITTI, Python

GPU-optimized PointPillars inference pipeline using TensorRT with parallelized preprocessing and postprocessing.

Technical focus: Real-time perception optimization with quantitative speed-versus-accuracy trade-off analysis.

• Achieved ~31-34x end-to-end speedup over serial baselines across tested range-image settings.
• Maintained comparable detection quality with a small accuracy trade-off under aggressive optimization.

Stack: Python, C++, CUDA, TensorRT, OpenPCDet, KITTI

Closed-loop perception + control integration in CARLA for sign-aware autonomous vehicle behavior.

Technical focus: Closed-loop linkage from vision outputs to control actions, with end-to-end simulation evaluation.

• Measured end-to-end perception-to-control response timing in closed-loop simulation.
• Evaluated sign-detection quality across scenario-driven test runs.

Stack: CARLA, TensorFlow, Computer Vision, Python

Kernel-level scheduling and demand-paging work with starvation prevention and page-fault handling.

Technical focus: Low-level systems work in scheduler fairness, virtual memory management, and kernel-boundary debugging.

• Validated scheduler fairness with stress tests across mixed CPU-bound and I/O-bound workloads.
• Verified page-fault handling behavior across demand-paging memory scenarios.

Stack: C, Xinu, QEMU, GDB