Abhijeet Gupta. — ML Engineer & Quant Researcher

Reasoning Model Failure Analysis, LLM Interpretability

A controlled LLM evaluation pipeline spanning six reasoning models from 7B to 70B parameters, designed to disentangle reasoning length effects from forced re-entry interventions. The study measured a 36-point accuracy decline in Llama-distilled models while Qwen-distilled models remained robust. Multi-GPU inference was conducted with a bfloat16 KV cache on 4x GH200 GPUs.

github.com/abhijeetgupta02/reasoning-model-failure-analysis-companion

Adversarial Robustness via Entropy Based Feature Selection in RL

An entropy-based feature selection framework for reinforcement learning agents that achieved 94 and 95 percent accuracy on Lunar Lander and Bipedal Walker under adversarial perturbations, outperforming KL Divergence and Joint Entropy baselines across Gym environments. The public reconstruction now ships standalone entropy and mutual-information diagnostics with adaptive histogram binning and a majority-class baseline.

github.com/abhijeetgupta02/adversarial-rl-feature-selection

Mouse Brain Cell Segmentation in Fluorescence Microscopy

A deep learning segmentation pipeline for high-noise fluorescence microscopy images, comprising a CNN architecture and a custom preprocessing routine for automated cell boundary detection.

github.com/abhijeetgupta02/mouse-brain-cell-segmentation

Virtual Yoga Instructor with Real Time Feedback

A real-time pose estimation and corrective feedback system using normalized joint angle features and repetition counting. The system achieves sub-100 ms latency on standard hardware and remains robust to variations in body size and camera angle.

github.com/abhijeetgupta02/virtual-yoga-instructor

VLM Evaluation Harness, Multimodal Model CLI

A command-line evaluation harness for vision-language models that emits structured JSON and CSV logs per run. The tool is designed for reproducible multimodal benchmarking with explicit configuration, deterministic seeds, and per-sample traceability. Metrics now include exact-match and SQuAD-style token-F1 with per-example F1 traceability and a constant-baseline reference.

github.com/abhijeetgupta02/vlm-eval-harness

Robust RL Observation Noise Benchmark

A CartPole observation-noise benchmark with a DQN baseline and extensible noise processes, designed for controlled evaluation of reinforcement learning robustness under sensor perturbations. The suite now covers the evaluation loop, DQN agent, and noise processes, with input-validation hardening.

github.com/abhijeetgupta02/robust-rl-observation-noise-benchmark

AI for Construction Safety, Evidence Package

A construction-safety vision evaluation reaching F1 0.7591, packaged with paired statistical tests, decision-threshold sensitivity analysis, and end-to-end claim traceability from raw predictions to reported numbers. Sparse-category rates are now reported with Wilson, Agresti-Coull, and exact Clopper-Pearson confidence intervals.

github.com/abhijeetgupta02/ai-for-construction-safety

Privacy-Preserving Career Prediction Benchmark

A privacy-preserving multi-output career-prediction benchmark built on synthetic data, paired with an entropy-based feature selection routine that controls disclosure while preserving downstream predictive utility. The benchmark now includes McNemar and Cochran-Q significance tests, Cohen's kappa agreement, and per-class diagnostics.

github.com/abhijeetgupta02/career-prediction-benchmark

Options Pricing Engine and Greeks Computation

Black-Scholes closed-form and Monte Carlo pricers with 100K antithetic paths for European equity options. Delta, Gamma, and Vega are computed both analytically and via finite differences across strike and maturity grids. Sensitivities now extend to first- through third-order Greeks (vanna, vomma, charm, speed, zomma, color), each cross-checked against finite differences.

github.com/abhijeetgupta02/options-pricing-and-greeks

Statistical Pairs Trading Backtest

An Engle-Granger market-neutral strategy applied to AAPL/MSFT, KO/PEP, and XOM/CVX over daily data from 2015 to 2023. The backtest incorporates walk-forward cointegration screening, out-of-sample hedge ratios, and a 1 bp transaction cost assumption. Sharpe ratio, maximum drawdown, turnover, and signal decay are reported. Mean-reversion diagnostics now include the Hurst exponent, a Lo-MacKinlay variance-ratio test, and a Kalman time-varying hedge ratio.

github.com/abhijeetgupta02/statistical-pairs-trading

GARCH Volatility Modeling and Stochastic Time Series

A GARCH(1,1) implementation applied to SPY and five single-name equities, validated using AIC, BIC, and Ljung-Box diagnostics. GARCH, LSTM, and rolling volatility baselines were benchmarked across the 2020 and 2022 stress periods through walk-forward error analysis. Forecast evaluation now adds Mincer-Zarnowitz calibration, HAC (Newey-West) Diebold-Mariano tests, and range-based Parkinson/Garman-Klass volatility estimators.

github.com/abhijeetgupta02/garch-volatility-modeling

LSTM Based Financial Time Series Forecasting

A stacked LSTM trained on over ten years of daily equity price and volume data, evaluated out-of-sample on the 2022 to 2023 period against ARIMA, GARCH, and random walk baselines using walk-forward error decomposition. Forecast comparisons now report Diebold-Mariano significance (with a Newey-West variance option) and the Theil U2 statistic.

github.com/abhijeetgupta02/financial-time-series-forecasting