Matthew Poncini

• Fine-tuned LLaMA-2-7B on a domain-specific financial question-answering dataset using PyTorch, Hugging Face Transformers, PEFT, TRL, BitsAndBytes, and QLoRA.
• Built a complete training and evaluation pipeline with separate training and held-out test sets, covering tokenization, batching, GPU execution, generation, output extraction, and metric computation.
• Used 4-bit NF4 quantization and parameter-efficient LoRA adapters while keeping base-model weights frozen.
• Improved ROUGE-1 from 0.1205 to 0.2519 and ROUGE-Lsum from 0.0818 to 0.1526 after fine-tuning.
• Evaluated hallucination, truncation, verbosity, domain adaptation, and behavioral overfitting in addition to aggregate metrics.

• 

  ROUGE before and after fine-tuning, charted from the scores reported in the paper.

• Designed an end-to-end ML experiment testing whether economic-news language contained useful information for predicting subsequent foreign-exchange movement.
• Collected and labeled 1,160 historical news observations, aligned them with future EURUSD, EURJPY, and USDJPY price changes, and created directional classification targets.
• Fine-tuned three specialized LLaMA-2-7B QLoRA adapters and evaluated them against held-out examples.
• Identified model collapse toward the majority Neutral class despite improving aggregate accuracy, tracing the failure to class imbalance and insufficient discriminatory signal.
• Proposed revised sampling, weighting, feature, and objective strategies based on observed model behavior rather than treating benchmark improvement alone as success.

• 

  Directional accuracy before and after fine-tuning, by currency pair.

• 

  Reference label distribution: Neutral dominates all three pairs.

• 

  On rising-price windows the tuned model scored 0%, the signature of collapse onto Neutral.

• Built Python experiments using five independently evolving genetic-algorithm populations with different parameter configurations to solve open-shop scheduling problems.
• Designed fitness functions, selection and mutation workflows, experiment tracking, parameter comparison, and statistical evaluation.
• Aggregated independently optimized schedules through a Wisdom of Crowds approach and evaluated the effect on solution quality.

• 

  Resulting schedule for 15 jobs across 10 machines, makespan 4591.

• 

  A 10-job, 10-machine schedule with makespan 370.

• 

  Each generation blends clones, experts, and three Wisdom of Crowds consensus crossovers.

Genetics and Wisdom of Crowds Hybrid Algorithm for the Traveling Salesman Problem

Folded a Wisdom of Crowds aggregation step directly into the crossover operator of a genetic algorithm, so each offspring inherits high-frequency edges from the wider population as well as from its parents. Produced higher-quality tours than the prior GA-only implementation.

• 

  Min, average, and max tour distance converging over generations.

• 

  A completed offspring tour over the city set.

• 

  Average distance by operator: aggregate, crossover, and mutation variants.

Unsupervised Clustering and Pattern Recognition of a 9-mer Peptide Dataset

Clustered one million 9-mer peptides drawn from human, viral, bacterial, and cancer sources using HDBSCAN and UMAP. Found clear physicochemical structure, but none of the 229 engineered features could reliably recover a peptide's biological origin.

• 

  UMAP + HDBSCAN recovers three clear structural clusters.

• 

  The same projection colored by biological source shows no separation, which was the negative result.

• 

  Sub-clustering within each top-level cluster.

Classification of a Trojan Horse Data Set

Built and compared six binary classifiers over ~160k web-traffic flows with 85 features, then stacked the top performers into an ensemble to separate trojan from benign traffic.

• 

  Random Forest ROC at AUC 0.767, the strongest single classifier.

• 

  Ensemble stacking ROC, AUC 0.722.

• 

  PCA cumulative variance across the 85 flow features.

Simulated Annealing and Niching Genetic Algorithms for the Bottleneck Traveling Salesman Problem

Surveyed exact and approximate approaches to the bottleneck TSP, then compared simulated annealing against niching genetic algorithms on an objective whose landscape is dominated by plateaus.

• 

  Average bottleneck edge length against problem size for every algorithm tested.

• 

  Runtime against problem size, log scale.

• 

  Bottleneck distribution by algorithm at 100 cities.

Niching Genetic Algorithm Experimentation

Applied a standard GA, deterministic crowding, and parallel hillclimbing to multimodal benchmark functions, analyzing convergence, population distribution, and each method's ability to hold multiple optima.

• 

  Baseline GA collapses its whole population onto a single peak of M1.

• 

  Deterministic crowding holds population across all five peaks.

• 

  Fitness convergence under deterministic crowding.

Gradient Descent Optimization Methods

Implemented and compared vanilla gradient descent, Newton's method, AdaGrad, and Adam across three-dimensional surfaces at varying starting points and learning rates.

• 

  Descent path across the Rosenbrock banana valley.

• 

  Descent path on the bowl function.

• 

  Adam: 3D path, contour path, convergence, and parameter evolution.

Trie Data Structure for Automatic Search Completion

Designed a trie-backed autocomplete structure and evaluated its lookup and insertion behavior against the demands of interactive search.

• 

  The prefix ban expands to every stored completion.

• 

  grap returns grape, grapes, and graph.

• 

  A misspelled query resolves to the nearest stored word.

Round Robin CPU Scheduling

Implemented a round-robin CPU scheduler in C++ and analyzed how quantum size drives turnaround and waiting time.

• 

  Round-robin timeline: three jobs cycling through 5 ms quanta.

• 

  Scheduler trace and wait-time summary.

• 

  1,001 jobs completed, average wait 28.8 s.

• Led technical personnel supporting more than $10 million in precision test and calibration equipment; diagnosed complex system failures through structured testing, validation, troubleshooting, and root-cause analysis.
• Managed technical workflows, documentation, accountability, and competing operational priorities in a high-reliability environment.

• Taught mathematics, statistics, programming, and quantitative reasoning while translating complex technical concepts for users with different levels of technical experience.
• Designed curricula, tools, assessments, and workflows around user needs and iterated based on direct feedback and observed outcomes.

• Supported pricing and retention research using SAS, statistical analysis, quantitative modeling, data validation, and business-impact analysis.

Relevant study: generative AI, artificial intelligence, machine learning, optimization, algorithms, data mining, databases, software engineering, computer systems, applied statistics, and C++ programming.

Relevant study: probability, statistics, regression, numerical methods, mathematical modeling, optimization, economics, and game theory.