I'm

Avrodeep Pal

★

Security & Infrastructure Hardening: Led security audits across a large-scale Python codebase, identifying and fixing critical path traversal and error handling vulnerabilities that affected all worker hosts. Redesigned the error relay system to return proper HTTP status codes, restoring monitoring accuracy and retry logic across the platform. Resolved production configuration drift where workers ran stale AI engine code, implementing self-healing deployment scripts and restoring metrics flow into usage records with 80+ regression assertions.

File Transfer Architecture Redesign: Overhauled the upload/download pipeline to eliminate browser hangs and memory spikes, achieving 470x reduction in RAM usage for large file uploads through streaming multipart handling and temp-file spooling with integrity verification. Built frontend concurrency controls and cache batching that reduced localStorage operations by 200x, eliminating platform freezes during multi-file operations. Implemented async thread pooling for git operations, enabling large repository syncs without blocking the event loop.

AI Model Benchmarking & Deployment: Designed and executed a hardware benchmarking framework evaluating 15+ vision and text models across laptop, cloud GPU, and server-grade hardware, documenting speedup factors up to 325x across 15 configurations. Set up and fine-tuned architectures for vision, ASR, and TTS models on server GPUs, building an ensemble of compact models that achieved 84% accuracy on skin lesion classification - benchmarked against a 400M-parameter foundation model that it outperformed by 1.5 points. Deployed these as multi-modal use cases directly on the platform.

Generic Workflow Architecture: Led the upgrade from custom-only nodes to a reusable workflow system capable of recreating domain-specific solutions (legal, financial, credit, medical) using generic building blocks - vector DB retrieval, LLM generation, and dropdown-driven prompt builders. Demonstrated this by rebuilding the Legal Helper workflow with generic nodes, reducing new workflow development time by 50% while expanding no-code accessibility for non-technical users. Created documentation and tutorials showcasing local model access without API restrictions.