⌨️ Open Source

Building a Qwen-style 0.6B transformer from scratch in PyTorch, end-to-end pretraining recipe with tokenizer, data pipeline, and training loop.

Continual pretraining recipes for adapting base LLMs to new domains without catastrophic forgetting. Includes data mixing strategies and eval pipeline.

LoRA / QLoRA fine-tuning recipes for Gemma3, optimized for single-GPU and multi-GPU setups. Includes data formatting and inference scripts.

Temporal Attractor Steering: a retrieval-free, training-free inference-time framework that detects, localizes, and steers parametric temporal conflicts in open-weight LMs. Project page with the verified PTC benchmark and code.

Structured runtime safety monitor for tool-using LLM agents. Deterministic rules + calibrated risk scoring under a four-class intervention policy (allow / block / confirm / revise). Project page with datasets and code.

Retrieval-augmented agentic AI system combining vector search with multi-step reasoning loops. MIT-licensed reference implementation.

Chat with multiple PDFs at once via a RAG pipeline. Vector store, chunking strategies, and a clean conversational interface.

A small domain-adapted language model for healthcare and clinical tasks. Pretraining + instruction tuning recipe on medical corpora.

Deep learning pipelines for detecting fractures on X-ray imagery. Compares ResNet, DenseNet, and EfficientNet backbones on benchmark data.

Multi-omics machine learning pipeline for RNA-Seq data analysis. Companion code for biological reasoning research.

Benchmark for schema-constrained extraction from nanobubble and nanocarrier literature. 51,566 deduplicated records with 40 gold-annotated examples across 18 fields (size, zeta potential, stability, loading efficiency, release profile).

1,874 biomedical research abstracts labeled by cancer type (Lung, Thyroid, Colon, Generic). For text classification and biomedical NLP research.

11.9k biomedical claim pairs from scientific articles for NLI, claim verification, and contradiction detection. Includes labeled contradictions, evidence spans, and publication metadata with train/val/test splits.

Temporal QA benchmark with 9,250 examples testing how LLMs handle knowledge conflicts and reasoning over time-varying facts. Wikidata-sourced with old/new entity values and their time periods.

Interactive neural network visualizer for teaching forward passes, activations, and gradients. Useful for explainability and intro courses.

Quantum machine learning experiments and notebooks. Comparing classical and quantum approaches on small benchmark tasks.

DeepSpeech2-based speech recognition implementation from earlier industry work. CNN + RNN architecture for end-to-end ASR.