Saumik Dana

I continue to develop and operate systems built during my time at Zebra, now running on my own capital. I switched the LLM-backed intraday options system to NVIDIA Nemotron 30B MoE and added an automated daily evaluation harness measuring reliability, rule compliance, consistency, and robustness.

I consolidated TradeStation market-data collectors into a reusable Terraform module and built a Kubernetes framework for multi-asset interday options backtests with indexed job scheduling.

I added a news-relevance pipeline to the intraday FX system using BM25, Qwen3 Embedding, reciprocal rank fusion, and Qwen3 Reranker; scaled LLM fine-tuning to single-node multi-GPU DDP; and added GEPA prompt optimization with gpt-oss-120b as the reflection model.

I deployed multi-asset, multi-strategy interday equity-options systems and built a scalable, memory-efficient Python backtesting framework supporting multi-asset strategies, signal-based execution, dynamic contract selection, delta hedging, and comprehensive P&L analytics.

I fine-tuned Amazon Chronos-T5 with LoRA on labels derived from the backtesting engine, then used the resulting embeddings to train per-asset TabPFN classifiers. I also built a multi-step, tool-calling signal-mining agent using gpt-oss-120b and MCP, later replaced it with a fixed protocol, and evaluated candidates with multiple-testing diagnostics to filter spurious signals.

I deployed multi-asset, multi-strategy intraday equity-options systems. I calibrated a jump-, rate-, and dividend-aware stochastic-volatility model to live option-chain snapshots and implemented Bayesian anomaly detection over its forecasts.

The deployed systems achieved 5–10% weekly returns on capital at risk. Their LLM-backed decision layer used deterministic entry gates and fed implied-volatility-skew narratives to gpt-oss-120b for hold and exit decisions.

I deployed a QLoRA-tuned Llama-3.1-8B-backed intraday FX trading system. gpt-oss-20b compressed news and price action into the inference context.

I fine-tuned a frozen NF4 base with rank-16 LoRA, evaluated held-out accuracy and precision for base and adapted policies using vLLM continuous batching, and built an OpenAI-compatible vLLM endpoint with versioned adapter promotion and storage-backed rollback.

I built a daily rebalanced, regime-aware equity portfolio using a Hidden Markov Model for regime detection, ElasticNet forecasts, and NSGA-II optimization, achieving 5% annualized alpha in a multi-year backtest. I also built a multimodal RAG assistant using CLIP embeddings, Qdrant retrieval, a vision LLM with a two-pass self-critique guardrail and page-level citations, a TypeScript/Next.js frontend, and a FastAPI backend on Modal.