Devansh Bansal

Python • LightGBM • CatBoost • scikit-learn • pandas

• Built end-to-end quantitative trading pipeline: 22-feature cross-sectional model (momentum, volatility, mean reversion, microstructure) across 100 anonymized equities over 10 years of daily OHLCV data
• Engineered rolling ensemble of Ridge Regression + LightGBM + CatBoost with 252-day adaptive retraining, achieving 410% cumulative return vs. 320% benchmark
• Implemented grid-search hyperparameter optimization over signal weights, smoothing alpha, and quantile thresholds to maximize Sharpe ratio with 10 bps transaction cost modeling
• Designed inverse-volatility position sizing and rank-transformed regression targets to reduce sensitivity to market regime shifts and extreme outliers

Python • SciPy • scikit-learn • NumPy • Matplotlib

• Credit Risk Modeling: Trained Logistic Regression and Random Forest classifiers achieving 99.9% accuracy and 0.9999 ROC-AUC on loan default prediction; implemented expected loss framework (PD × EAD × LGD)
• FICO Bucketing: Developed optimal credit segmentation via constrained non-linear optimization (L-BFGS-B), comparing MSE vs. log-likelihood objectives to partition continuous scores into risk tiers
• Commodity Forecasting: Built time series model decomposing natural gas prices into trend and seasonal components using harmonic sine/cosine feature engineering, achieving R² > 0.85
• Storage Contract Valuation: Engineered deterministic DCF framework pricing gas storage contracts across seasonal spreads with injection/withdrawal mechanics and inventory constraint validation

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

• Architected 3-phase NLP pipeline: audio ingestion (yt-dlp + ffmpeg), multilingual ASR (Whisper large-v3), and LLM-based concept extraction with structured JSON output and DAG validation
• Implemented graph normalization with cycle enforcement using confidence-ranked greedy edge insertion and topological sorting for pedagogical flow ordering
• Built robust LLM output parser handling code-mixed Hindi-English transcripts with concept aliasing, ID sanitization, and flexible JSON extraction

C++ • Graph Theory

• Implemented and benchmarked Ford–Fulkerson, Dinic's, and Push–Relabel algorithms; analyzed asymptotic complexity on dense, sparse, and adversarial graph structures

C • Sockets • Pthreads

• Built multi-threaded networked file system supporting concurrent multi-user access with strict reader-writer locking and socket-based IPC