Multimodal Document Parsing [Python]
- Lead the effort (architect, design, implementation) on GenAI team's multimodal document parsing, used by over 85% of GenAI customers.
- Supports multiple document types (text, images, tables, audio, video).
- Improve qualities with various features of documents (hierarchical extraction, page stitching etc), as well as automatic tuning techniques with LLM.
- Designed an async architecture that improves speed by over 10x.
Programming Language Translation (Fortran to C++) [C++, Python]
- Using LLM with RAG for low-level unit translation, augmented with syntax-aware, language agnostic source code parsing.
- Test on the original tests with C FFI (> 99% accuracy).
- Wrote an UI that handles LLM calls and compilation in an async manner.
RL Chip Design [Python, C++]
- Designed and implemented an RL AI / EDA system speeding up initial chip development of the world's biggest chip design by 3x.
- Proposed and validated RL agent-based chip design models to justify solution quality.
Multihop Retrieval Reasoning Model [Python]
- Improved a RAG system and reduced error rate by 40% with multihop document retrieval.
- Speed up the distributed LLM trainer by 30% via careful profiling and critical-path reduction.
MLE / performance improvement [Python]
- Integrated ML (vector retrieval) pipeline functionality into pandas pipeline.
- Interactive debugging capabilities which found a critical SQL bug in the first week of deployment.
10x faster LLM evaluation with Bayesian optimization and NLP
[Python]
- This is an automated system that speeds up LLM benchmarking by 10x.
- Uses dense retrieval on embeddings to quantify the queries in the corpus
- Then uses bayesian optimization to select the best queries.
ML library for fixing CUDA (GPU) issues in single line
[Python]
- Uses lazy evaluation to solve out-of-memory errors in PyTorch with a minimal API.
- Supports CNN, GNN, RNN, Linear layers, and arbitrary PyTorch ops.
ML book in question answer format (StackOverflow style)
[Python]
- Compile from, and serve to, 1300+ students (ML class where I was a TA).
Stock allocation advisor powered by explainable AI LLM agents [Python]
- Led a team of 4 to win 1st place out of 1500 participants in BlackRock's event.
- Designed fintech ML models for transparent NLP-based financial decisions.
- Built backend for volatility index and portfolio aggregation.