AI Evaluation · Benchmark Design · Structured Reasoning · Retrieval Systems

I focus on AI evaluation and benchmark design — building structured frameworks that produce honest, reproducible results. My work emphasizes evaluation methodology: leakage-free design, criterion-level reasoning, error analysis, and transparent reporting of limitations and negative results. I have hands-on experience in LLM output evaluation, annotation QC, and dataset validation workflows.

My background is in Biomedical Engineering, which is where my projects are grounded — but the skills are methodological, not domain-specific.

• Designing leakage-free ML evaluation pipelines with reproducible methodology
• Structured LLM output evaluation: criterion-level reasoning, error analysis, failure mode documentation
• Annotation QC and dataset validation workflows
• Hybrid information retrieval with graded evaluation (BM25, semantic embeddings, hybrid ranking)
• Subject-independent cross-validation for time-series classification

Structured benchmark for LLM-assisted eligibility reasoning across 150 synthetic patient–trial pairs.

• Criterion-level reasoning with robustness checks, error analysis, and an HTML benchmark report
• LLM-reviewed draft labels; results reported with known limitations (Accuracy 0.687 · Macro F1 0.686)
• Fully synthetic patients; not clinical decision support

Leakage-aware ML benchmark for wearable IMU classification, built with strict evaluation hygiene.

• Normalization fitted on training folds only; subject-independent 5-fold cross-validation
• Results with confidence intervals: balanced accuracy 0.6441 ± 0.0156, AUROC 0.7029 ± 0.0261
• Designed for realistic evaluation, not inflated scores

FastAPI + Streamlit search system with a hybrid retrieval pipeline and full graded evaluation.

• BM25, semantic embeddings, and hybrid ranking with optional filters
• Evaluated with Precision@5, Recall@10, MRR, and nDCG@10; best hybrid result: MRR 1.0000 · nDCG@10 0.9453

Evaluation & Benchmarking Leakage-aware evaluation design · Structured output evaluation · Error analysis · Annotation QC · Dataset validation · Limitation documentation · Subject-independent cross-validation

Retrieval & ML BM25 · Semantic embeddings · Hybrid ranking · Classical ML · 1D-CNN · Feature engineering · Threshold analysis

Languages & Libraries Python · NumPy · Pandas · Scikit-learn · PyTorch · SciPy · FastAPI · Streamlit

Tooling Git · Jupyter · SQL · REST APIs

• GitHub: github.com/gabrielepecchi