Hi, I'm M Faisal Chughtai
Full Stack · Data Engineer · ML

Developing a machine learning framework to predict valuations for TCGs like Pokémon and Magic: The Gathering. Modelling intrinsic factors (age, condition) and extrinsic drivers (playability, influencer trends, grading populations) using regression and time-series analysis.

Scraping historical price data to assess volatility and reliability of future set valuations. Investigating the impact of third-party grading (PSA/BGS) as a measure of demand and market efficiency in alternative investment spaces.

Proposed forecasting GB power system inertia using a Bi-LSTM model benchmarked against ARIMA and XGBoost, trained on NESO 30-minute settlement data across wind, solar, thermal, and interconnector sources. Addresses the growing stability risk as renewable generation displaces synchronous machines.

Designed a SHAP-based explainability layer to break down individual inertia predictions into per-feature contributions. Includes beeswarm plots, force plots, and time-of-day SHAP patterns to make model behaviour transparent for grid operators.

U1 built up from biological neurons to perceptrons, activation functions, and the universal approximation theorem. U2 went deeper into FFNs: vanishing gradients, Xavier/He initialisation, batch normalisation, dropout, and modern optimisers (Adam, SGD with momentum).

U3 covered CNNs, pooling, and key architectures (LeNet, VGG, ResNet) with transfer learning. U4 introduced RNNs, LSTMs, GRUs, and seq2seq. U5 tackled the Transformer: self-attention, positional encoding, BERT pre-training, GPT generation, and LLM fine-tuning trade-offs.

Built a full data mining pipeline for epitope prediction on Trypanosoma cruzi peptide sequences (binary classification). Performed EDA on class-imbalanced data, applied feature scaling and SMOTE, then used GroupKFold cross-validation with the Info_group column to prevent data leakage across protein families.

Trained and compared Random Forest, SVM, Logistic Regression, and Gradient Boosting classifiers using balanced accuracy as the primary metric. Selected the best model, generated predictions on an unseen holdout set, and submitted results as a reproducible notebook, PDF report, and predictions CSV.

Implemented a Battleships/Yubotu solver in Python on a 10x10 grid. Built ship placement validation, hit/miss tracking, win condition logic, and an AI guessing strategy using checkerboard search patterns combined with targeted hunt mode after a confirmed hit.

Full data science pipeline on the VGSales dataset: cleaning, EDA, feature engineering (decade, regional share), and statistical testing (t-test, ANOVA). Built and compared Linear Regression and Random Forest models for global sales prediction. Submitted SQL queries separately covering aggregation, filtering, and window functions.

Investigated a 5-feature dataset to determine classification vs. regression task type and assessed linear vs. non-linear model fit with evidence. Written a technical email to a colleague explaining learning rate in MLP context, its effect on convergence and generalisation, and practical guidance on selection without prior dataset knowledge.

Applied k-means and GMM with principled cluster selection to a 450-point dataset. Used Sparse PCA on New Delhi air quality data before and after standardisation, comparing component sparsity. Designed and compared 3+ classifiers for osteoporosis risk prediction (5,000 patients), integrating medical literature on Vitamin D, Calcium, and alcohol thresholds into preprocessing decisions.

Community detection on a real European email network using the Girvan-Newman algorithm. Computed modularity scores, degree distributions, and clustering coefficients to identify organisational clusters.

Epidemic modelling on a 6,486-node Marvel character network. Computed the critical spreading threshold (λc ≈ 0.78%) via Molloy-Reed, and ran stochastic SIR simulations to analyse phase transitions from local outbreaks to full epidemics.

Analytically derived MLEs for Exponential and Rectified Gaussian distributions on 400 call-centre records, selecting the best fit to flag outliers. Extended to a Bayesian gemstone classifier, deriving the optimal decision threshold mt analytically, including asymptotic and asymmetric cost cases.

Fitted GMMs (K = 3, 4, 5) to penguin culmen-depth and weight data via EM. Plotted 2D distributions and log-likelihood convergence per K, then used BIC/AIC to select the optimal number of species and discussed why results may diverge from ornithologists' expectation of 3.