Building reliable software.
Making AI useful.

AI Research Intern

• Developed and evaluated an initial machine-learning approach to predict bispecific-antibody manufacturing yield from sequence data. Limited labeled training data and very low, inconsistent evaluation performance prevented reliable predictions, motivating a shift toward a more measurable ML problem.
• Helped redefine the research objective as sequence-pair matching and ranking: predicting which two protein chains naturally belong together. This enabled large-scale supervised learning and more rigorous comparisons against established models.
• Engineered large-scale Python data pipelines to preprocess, annotate, clean, and validate approximately 3.75 million paired sequences. Implemented metadata enrichment, dataset filtering, negative-sample construction, and data-integrity checks to support reproducible training and evaluation.
• Built and optimized PyTorch training and evaluation workflows, using pretrained sequence embeddings and reusable feature caches to process millions of examples efficiently. Experimented with classification, ranking objectives, and negative-sampling strategies to improve model performance.
• Conducted extensive model evaluation, statistical benchmarking, and error analysis, examining dataset biases, sampling effects, and unreliable labels. On a matched held-out benchmark, achieved 0.798 ROC-AUC (κ) and 0.823 (λ) versus 0.750 and 0.783 for ImmunoMatch, respectively.
• Benchmarked against p-AbNatiV2 on identical held-out candidate sets, achieving 79.66% vs 74.49% accuracy in 1-vs-1 matching and 20.72% vs 8.56% Top-1 retrieval accuracy in 1-vs-50 ranking (κ), while identifying limitations requiring further research.