Data scientist with 5+ years shipping production pipelines, inventory systems & GenAI applications. I own the full arc — scoping, building, and landing the "so what" with leadership.
Most of my five years has been in retail and commerce: at Wayfair, owning end-to-end data pipelines for inventory allocation and replenishment across 10K+ SKUs; at Pantone, leading sales and marketing analytics — LTV, CAC, cohort ROI — that translated into pricing and GTM moves.
Lately I've been pulling that same discipline into GenAI. I built a Slack-based SQL bot at Wayfair that put self-serve analytics in stakeholders' hands, and I've been shipping a portfolio of production-grade GenAI systems — RAG with evaluation gates, local LLM benchmarking, QLoRA fine-tuning — applying the same correctness-and-measurement instincts that pipelines taught me.
The through-line: I care about building systems that produce correct, auditable outputs — and proving it.
Two production-grade GenAI projects. Both run, both evaluated, both on GitHub.
Domain-specific document Q&A with hybrid retrieval, re-ranking, citation enforcement, and CI-integrated evaluation. Instrumented with per-stage latency tracing, cost-per-request tracking, and a Streamlit dashboard. A GPT-4o-as-judge harness blocks deploys on faithfulness regression.
A data-driven model-selection report measuring latency, throughput, and structured-output accuracy for Llama 3.2, Phi-4-mini, and Mistral 7B — with quantization tradeoff analysis and Pydantic-validated output with retry logic.
What's queued. Updated as things ship.
Fine-tuning Qwen2.5-3B-Instruct for structured JSON extraction from e-commerce product descriptions — chosen for retail overlap and clean field-level accuracy evaluation.
Extending the RAG-project eval approach into a standalone library: dataset generation, rubric scoring, cost tracking, and regression detection across model versions.
An agent answering operational questions over a simulated inventory warehouse using tool use, SQL generation, and self-critique. Bridges the portfolio back to the domain I know best.
Five years of assembled tools — from warehouses and pipelines to GenAI backends.
Short-form writing on what I'm building — the decisions, the surprises, the measured deltas. No LinkedIn thinkpieces.
Open to data-science and AI-engineering roles globally. Happy to chat about retail systems, retrieval, or what broke on your last LLM launch.