Zebra Capital Management LLC · Stamford, CT
1.5+
Sharpe, agent-mined strategies
5–10%
Weekly returns on capital at risk
5%
Annualized alpha vs. benchmark
LLM-backed trading systems
- Fed risk-neutral volatility, jump-risk, and repricing signals from live option-chain calibrations to gpt-oss-120b for intraday entry and exit decisions.
- Built an automated LLM evaluation harness grading every snapshot for reliability and compliance, plus daily consistency and robustness tests.
- Deployed an intraday FX system feeding news and price action to a DPO-QLoRA fine-tuned Llama-3.1-8B with GEPA-optimized prompts, executed through OANDA.
- Replayed archived market contexts against ground truth to benchmark precision, recall, F1, coverage, and selective accuracy.
- Trained a rank-16 BF16 LoRA over a frozen 4-bit NF4 base in TRL/PyTorch, accelerated preference generation with vLLM continuous batching, and served the promoted policy on on-demand L4 GPUs.
- Ran evaluation-gated prompt optimization (DSPy/GEPA) with gpt-oss-120b as the reflection model—a new prompt ships only when it beats the incumbent.
- Deployed an intraday commodity system on MiniMax-M2, backed by hybrid retrieval: BM25, Qwen3 Embedding, reciprocal rank fusion, and a Qwen3 Reranker.
ML-backed trading systems
- Deployed intraday options strategies on Bayesian anomaly detection over stochastic-volatility implied vols—5–10% weekly returns on capital at risk.
- Ran a daily rebalanced, regime-aware equity portfolio: HMM regime detection, ElasticNet forecasts, NSGA-II optimization—5% annualized alpha in a multi-year backtest.
- Deployed interday options trading on trained TabPFN classifiers, backed by a multi-asset, multi-strategy backtesting engine.
- Fine-tuned Chronos-T5 with LoRA on backtest-derived labels, then trained per-asset TabPFN classifiers on the resulting embeddings.
Agents and infrastructure
- Built a tool-calling LLM workflow that mined interday options entry signals with Sharpe above 1.5.
- Shipped an agentic multimodal RAG PDF assistant: CLIP embeddings in Qdrant, cosine and MMR retrieval, cited answers from a two-pass Qwen3 vision model.
- Deployed containerized serverless systems on AWS and Modal, wiring Fireworks AI inference to HMM, Heston, Chronos, TabPFN, Qdrant, DSPy, and broker APIs.
- Engineered DST-aware scheduling, persistent position state, model-artifact caching, monitoring dashboards, execution recovery, and fail-closed error handling.
