I work on AI systems and living ones.

Cell2Compound: Auditing a Drug-Response Benchmark

RDKit, Scanpy, PyTorch, scikit-learn, ChEMBL, ESM-2

• Built a compound-held-out drug-response pipeline over sci-Plex3 single-cell RNA-seq, streaming 2.5 GB of expression data in 82 s and raising usable cells per condition from 36 to 215.
• Measured replicate noise floors across four datasets and 76,397 LINCS L1000 conditions spanning 1,750 compounds and nine cell lines, establishing that 74% of a standard perturbation benchmark carries no measurable signal.
• Showed a 24-dimensional mechanism-class label matches a 2048-bit Morgan fingerprint across four datasets and four models, and that demonstrable skill exists only in the top signal-to-noise quartile.
• Retrained chemCPA (NeurIPS 2022) from scratch to avoid test leakage, patching two blocking defects; a five-seed paired comparison reversed a single-seed result favouring it, leaving the two models statistically indistinguishable.

Read the write-up Source

Prediction of Future Continuous Motion States from ECoG Recordings

Scikit-learn, Pandas, Matplotlib, Research

• Built a data pipeline to analyze ECoG data for correlations between neural signals and cursor movement.
• Achieved a max R-square of 49.3%, with processing latency reduced to 5 milliseconds using techniques like frequency filtering and PCA.
• Identified correlations between Brodmann areas and neural signals, enabling faster processing by targeting specific brain regions.

Adversarial Tweet Sentiment Analysis

PyTorch, HuggingFace, Transformers

• Performed sentiment analysis using SBERT from HuggingFace, reducing high-dimensional data to 3D space with PCA.
• Trained a logistic regression model, validating data compression without loss in prediction accuracy.
• Highlighted classification challenges with slang and Twitter-specific words, identifying model limitations.

Autonomous Ice Hockey Agent

Python, PyTorch

• Developed an autonomous agent to play ice hockey using image-based and state-based approaches.
• Achieved 85% accuracy in ball tracking and over 80% game success through policy optimization with the REINFORCE algorithm.
• Built a data pipeline for training and assembling datasets to optimize the agent's goal-scoring strategy.

Analyzing Dataset Artifacts using ELECTRA

PyTorch, HuggingFace, Transformers

• Developed an NLI model using ELECTRA, achieving 88.24% accuracy and improving predictions by correcting dataset artifacts.
• Designed an error analysis framework, categorizing issues to enhance semantic processing and model robustness.
• Conducted experimental fine-tuning on diverse datasets, improving model generalizability.