# Kavin Ravi

> Kavin Ravi's personal portfolio. Projects and research in language models, optimization, and quantitative finance.

Kavin is based in California and expects to graduate from Chapman University in December 2026 with a B.S. in Data Science and a B.S. in Business Administration. His main interests are rigorous optimization in machine learning and quantitative finance. Outside his technical work, he enjoys rock climbing, tennis, and origami.

## Pages

- [About, experience, interests, and contact](https://kavinravi.com/)
- [Projects](https://kavinravi.com/projects)
- [Research](https://kavinravi.com/research)
- [Current resume](https://kavinravi.com/resume): Stable link to the current resume document.

## Projects

- [Axiom Tilt Strategy](https://kavinravi.com/axiomtiltstrategy): A live long-only S&P 500 portfolio system built by Kavin Ravi. It combines fundamental, quantitative, and macroeconomic features with machine learning to inform stock selection and allocation. Portfolio constraints manage concentration and position sizing. Kavin developed the financial data processing, allocation workflow, and monitoring software used to operate and review the portfolio. [Dashboard](https://kavinravi.com/axiomtiltstrategy/dashboard): Password-protected holdings, performance, and allocation history for reviewing portfolio decisions over time.
- [Helix ML](https://kavinravi.com/projects/helix-ml): An open-source ML experiment workspace using native Codex or Claude Code subscriptions. Upload labeled data, compare independently measured trials, and download verified model exports. The hosted CPU preview runs private workspaces on Railway and requires no local installations. [Open Helix](https://helix-web-production-03c9.up.railway.app). [Source](https://github.com/kavinravi/Helix-ML). MIT licensed.
- [SLM](https://kavinravi.com/projects/slm): A 538M-parameter Transformer with SFT/DPO post-training and a separate 535M-parameter Mamba2 model without post-training.
- [pytorch-opt](https://kavinravi.com/projects/pytorch-opt): Second-order and structure-aware PyTorch optimizers, shared numerical tools, diagnostics, and an optional C++/ATen backend.
- [Finance Dashboard](https://kavinravi.com/projects/finance-dashboard): Streamlit market analysis and forecasting with ARIMA and PyTorch LSTMs. Source: https://github.com/kavinravi/finance-dashboard
- [PsychAI](https://kavinravi.com/projects/psychai): Qwen3-8B LoRA adaptation using counselor Q&A. Research project, not a clinically validated service. Adapter: https://huggingface.co/kavin-ravi/qwen3-8b-psychai-lora

## Research

- [Axiom Tilt](https://kavinravi.com/research/axiom-tilt): Preprint by Kavin Ravi and Dylan Massaro, Chapman University. SEC filing features from continued-pretrained FinBERT feed a LightGBM stock ranker, followed by a PPO allocator. Text features improved all three allocators' after-cost Sharpe ratios against their no-text versions, although only the PPO comparison was individually statistically significant. PPO showed no reliable improvement over deterministic score-proportional allocation across three seeds; all three underperformed on the untouched 2025 holdout. The equity-only strategies showed no statistically significant risk-adjusted improvement over SPY. The study reports 2009–2024 results, comparisons excluding the 2009 configuration-selection year, and a 2025 holdout. Limitations include incomplete fundamentals coverage, simplified execution costs, and a base-versus-continued-pretrained FinBERT comparison on only one walk. These are historical backtest results.
- [Axiom Tilt paper](https://kavinravi.com/papers/axiom-tilt.pdf): Preprint dated September 30, 2026, by Kavin Ravi and Dylan Massaro.
- [Mamba2 inference latency](https://kavinravi.com/research/mamba2-latency): Ongoing research led by Kavin Ravi, supported by Prof. Jonathan Hersh at Chapman University. Investigates time to first token, inference throughput, and financial classification against Transformers, FinBERT, and ModernBERT. No validated results or paper yet.
- [Optimizer efficiency](https://kavinravi.com/research/optimizer-efficiency): Ongoing independent research led by Kavin Ravi. Planned comparisons of 100M/300M Mamba2 and Transformer models using AdamW, Muon, Shampoo, and K-FAC. Approximately 40–50 configurations using identical training documents. Primary metric: wall-clock time to a working target of 2.5 validation loss on FineWeb-Edu holdout data. The target is not an achieved result.

## Experience

- Syon Capital Management, Summer Analyst: June–July 2025, San Francisco.
- Viallion, Finance Assistant: June–August 2025, Menlo Park.
- Syon Capital Management, Seasonal Intern: June–July 2024, San Francisco.

## Contact

- Email: kavinravi121@gmail.com
- Phone: +1 650 250 3752
- [GitHub](https://github.com/kavinravi)
- [LinkedIn](https://linkedin.com/in/kavin121/)