Devansh Bansal

Python • OpenAI Whisper • GPT-4 • NetworkX • yt-dlp • Pyvis

• Architected modular 3-phase pipeline: audio ingestion (yt-dlp + ffmpeg normalization to 16kHz mono WAV), multilingual speech-to-text (Whisper large-v3 with CUDA/CPU auto-detection), and LLM-based structured concept extraction
• Built robust LLM output parser handling code-mixed Hindi-English transcripts with concept aliasing, ID sanitization (regex normalization + collision detection), and flexible JSON extraction from varied response formats
• Implemented DAG validation engine: confidence-ranked greedy edge insertion maintaining acyclicity, topological sort for pedagogical flow ordering, and interactive Pyvis graph visualization with physics simulation
• Designed CLI with configurable Whisper models, mock-LLM dry-run mode, fail-fast debugging, per-video error tracking with UTC timestamps, and JSON Schema validation for all structured outputs

C • RISC-V • Kernel Development

• Extended the xv6 kernel (RISC-V) with custom CPU scheduling policies and additional system calls; developed kernel-level tests to evaluate scheduling fairness and context switch overhead

C • Sockets • Pthreads

• Built multi-threaded networked file system enabling concurrent multi-user access with strict reader-writer locking mechanisms, socket-based IPC, and fault-tolerant connection handling

C • Unix Syscalls • Process Control

• Implemented a custom Unix shell supporting piping, I/O redirection, background process execution, job control, and robust signal handling (SIGINT, SIGTSTP, SIGCHLD) using low-level system calls

SQL • Python • DBMS

• Designed relational schema modeling complex datasets with constraints, triggers, stored procedures, and optimized analytical queries for multi-table joins and aggregation workloads

C++ • Graph Theory • Algorithms

• Implemented and benchmarked Ford–Fulkerson, Dinic's, and Push–Relabel algorithms; profiled performance on dense, sparse, and adversarial graph inputs to validate theoretical complexity bounds

Python • LightGBM • CatBoost • scikit-learn

• Built quantitative trading pipeline with 22-feature cross-sectional model and rolling Ridge + LightGBM + CatBoost ensemble achieving 410% cumulative return vs. 320% benchmark over 100 equities
• Implemented point-in-time backtesting with transaction cost modeling, inverse-volatility position sizing, and grid-search Sharpe ratio optimization