Saratchandra Patnaik

Problem: Build a telecom-grade server that receives encrypted binary UDP packets at high throughput and processes them without dropping or reordering under concurrent load.

Approach: Producer-consumer architecture with a thread pool and mutex-locked queues. Producer threads read raw UDP datagrams off the socket; consumers decrypt and process binary payloads in parallel. Statistics are published via POSIX shared memory IPC for external monitoring.

Result: High packet ingestion at sustained throughput with ordered processing guarantees and zero data loss under concurrent load.

Problem: Run real-time face recognition on an edge device with no pip access — cloud-based inference introduced too much latency and the Greengrass Lambda environment had no package manager.

Approach: Packaged raw PyTorch MTCNN and FaceNet models into a custom deployment bundle that runs in a pip-free Lambda environment on Greengrass. Edge inference events stream asynchronously to the cloud via MQTT and SQS, fully decoupling edge processing from cloud consumption.

Result: Sub-second face recognition at the edge with a cloud-synchronized event stream and no internet dependency at inference time.

Problem: CPU-based 2D Gaussian filtering was the bottleneck for image processing workloads — single-threaded and unable to scale with image resolution.

Approach: Parallel CUDA kernel with shared memory tiling to eliminate redundant global memory reads, loop unrolling to maximize instruction throughput, and grid-stride loops to handle arbitrary image sizes. Validated output fidelity against CPU reference using PSNR and SSIM.

Result: 20× speedup over CPU — full-resolution frames processed in milliseconds instead of seconds.

Problem: Build a composable AI assistant for vehicle diagnostics, ATS resume optimization, and document Q&A — without locking the system to a single LLM provider.

Approach: Clean Architecture to decouple domain logic from model providers (OpenAI, Anthropic). ChromaDB handles vector retrieval for document grounding; GPT-4 Vision processes image inputs for diagnostics. Each capability is an isolated use case wired through a shared retrieval layer. React frontend, FastAPI backend.

Result: Fully swappable model backends with consistent retrieval quality across task types and a usable dashboard for diagnostics and resume analysis.