🎓 Senior @ UIUC — CS + Philosophy
🔬 AI Safety Researcher | Systems Builder | Open-Source Contributor
🌐 LinkedIn · Email
I combine AI safety research with systems engineering. My focus is on mechanistic interpretability, probing truth in embeddings, and making LLM reasoning faithful.
I’ve worked on scaling PyTorch pipelines for probing embeddings, steering generative models, and GPU/distributed systems optimizations — with industry experience at Thomson Reuters, CCC Intelligent Solutions, and AMD collaborations, plus startup work at Onespace.
My goal: design safer, interpretable, and more controllable AI systems.
- 🧩 Mechanistic Interpretability (UIUC) – Identified truth/ambiguity subspaces in LLaMA, Mistral, Gemma embeddings (80% probe accuracy).
- 🎙 Steering Whisper – Built PyTorch steering pipeline; aligned GPT-2 ↔ Whisper concepts across modalities.
- 📊 Cognitive Battery for LLMs – Designed 1,000-item psychometric evaluation estimating causal covariance and effect sizes.
| Project | Description | Tech |
|---|---|---|
| 🎙 Steering Whisper | Built a PyTorch steering pipeline for OpenAI Whisper to align cross-model concepts (Whisper ↔ GPT-2), exploring transfer of reasoning, tone, and emotion directions. | PyTorch · Transformers · PCA/Projections |
| 🔗 Causal CoT Ablation | Analyzed chain-of-thought reasoning by ablating specific steps and measuring logprob shifts, highlighting stepwise faithfulness and causal influence in reasoning chains. | PyTorch · Mechanistic Interpretability |
| 🔍 Distributed-Log-Queryer | Go-based grep service across 10 VMs; fault-tolerant log streaming with a SWIM failure detection system. | Go · RPC · Goroutines |
| 📖 ViT-Prisma Contribution | Docs & fixes for open-source vision interpretability library. | PyTorch · Vision Transformers |
| 🧪 Phase-Transition Interpretability | Induced a training phase change (orientation → orientation+color) to study how weights change as neurons become polysemantic. Tracked weight-angle, magnitude, and entropy as order parameters | PyTorch · NumPy · Matplotlib |
| ⚡ GPU-Convolution-Optimization | CUDA + Tensor Cores optimization, 70% speedup. | CUDA · WMMA · Nsight |
| 📝 Onespace | JSON annotation + LangChain RAG chatbot for legal docs. | React · Electron · LangChain |
| 🧩 xv6 Operating System | RISC-V OS kernel: preemptive scheduling, paging (SV39), VirtIO drivers, UART logging. | C · OS · RISC-V |
- Thomson Reuters (SWE Intern) – Built legal chatbots and optimized UX for the 100 top law firms worldwide
- CCC Intelligent Solutions (Data Science Intern) – Fine-tuned vision fraud detection on 40M claims, ↑ accuracy 20%.
- Onespace (Founding Engineer) – Secured $50k funding, shipped RAG + annotation platform.
- Gies Disruption Lab (Tech Lead) – Built deepfake detection (↑30%), deployed RAG over 500k docs.
- Faithfulness of chain-of-thought reasoning in LLMs
- Subspace and causal probing for embeddings
- Practical tools for robust deployment of generative AI
⭐️ Let’s collaborate on AI safety & interpretability research → reach me via email or X!