AI research scientist — I train models and ship the systems around them.
My work runs from the architecture up: implementing transformers and optimizers from scratch, designing multi-task and meta-learning methods, and then putting those models into applications where accuracy has to survive contact with messy data.
First author on three papers at Stanford University, both with the code public and the results reproducible.
Extending BERT with Multi-task and Meta-learning · Stanford CS 224N
A single BERT encoder serving sentiment, paraphrase, and semantic similarity at once.
Shared projected attention layers, Siamese sentence encoders with an early u−v interaction
term, and Proto-BERT for few-shot classification.
74.0% average test accuracy, +20.4% over baseline — top 5 on the class leaderboard.
Proto-BERT reaches 38.7% on a 5-way 5-shot out-of-domain meta-test.
The transformer encoder and AdamW optimizer are implemented from scratch — no transformers.
TLDChoiceNet: Quantitatively Choosing a Transfer Learning Dataset · Stanford CS 330 Predicting how well a transfer-learning dataset will work before spending the compute to fine-tune on it. Cut prediction MSE 5×, and designed an unsupervised class-correlation metric that explains fine-tune accuracy with R² = 0.974 from a single forward pass — no training required. Along the way: evidence that ImageNet-pretrained weights actively push dissimilar classes apart in latent space.
| Project | What it does |
|---|---|
| LangCrab | Observability for LLM agents, built on LangSmith |
| Fake review detection | Opinion-spam classification on Yelp — BERT, LightGBM, LSTM, CNN-LSTM and Naive Bayes benchmarked against each other. Fine-tuned BERT wins at 77.9% accuracy, 0.766 F1 |
| 3D image prediction | Volumetric prediction from image data |
Modelling — Transformers, BERT, meta-learning (prototypical networks, MAML-style few-shot), multi-task architectures, transfer learning, CNNs, LSTMs, gradient-boosted trees
Engineering — PyTorch, TensorFlow/Keras, Python, LangSmith, TensorBoard, scikit-learn, LightGBM
The part people skip — reading the paper, implementing the method from scratch, and checking the result against a reference before believing it
Working on LLM agent observability, and interested in where meta-learning meets foundation models — how systems adapt to new tasks and domains from very little data.
Open to research collaborations.