AI Deployment / Forward Deployed Engineer — LLM evaluation, data & agent workflows. Based in New Jersey, USA.
I turn ambiguous technical problems into working AI systems: scoping agent solutions directly with business clients, building retrieval and tool-calling workflows, and designing the evaluation and human-review loops that keep them trustworthy.
- 🛰️ AI Specialist, STEM & Data Intelligence @ SpaceX (Apr 2024 – present) — LLM reasoning frameworks for STEM / advanced-math data, large-scale data curation and auditing, quality benchmarks for model outputs.
- 🧠 Previously AI Data Scientist & STEM Expert @ Outlier and Data Scientist @ Shiptory.
- 🎓 M.S. Information Technology (Data Management & Analytics) · B.S. Mathematics.
- 📚 Before tech: eight years leading a mathematics department and authoring calculus & geometry textbooks — I like explaining complex things clearly.
Ordered by relevance to forward-deployed work. These are reference implementations on synthetic data only — no client code or client data. Three of them (retail-whatsapp-llm-agent, turkish-legal-rag-intake, llm-eval-harness) are inspired by client projects I delivered.
| Project | What it shows |
|---|---|
| enterprise-integration-agent | LLM agent over mock CRM + ticketing REST APIs: OAuth2 (client credentials, PKCE, refresh rotation), HMAC-signed idempotent webhooks, retries with backoff, PII-redacted audit log. |
| customer-data-onboarding-pipeline | Onboarding a customer's messy CSV drops + paginated REST API into DuckDB/Postgres: data contracts with quarantine, LLM-assisted column mapping with human review and a held-out eval, dbt-style models, DQ checks, idempotent re-runs, audit trail. |
| retail-whatsapp-llm-agent | WhatsApp retail agent: tool-calling over a product DB (price / promotion / stock), escalation to staff, mock WhatsApp Cloud API webhook, store reports. |
| turkish-legal-rag-intake | Turkish legal RAG with citation-grounded answers, structured client-intake agent, attorney review queue, rubric-based eval. |
| text-to-sql-analytics-agent | Schema-aware text-to-SQL over a dbt-style DuckDB warehouse: read-only SQL guard (allowlist, PII, LIMIT), self-correction, execution-accuracy eval. |
| llm-eval-harness | LLM evaluation harness: rubric + reference scoring, LLM-as-judge, pairwise preference labels, Cohen's kappa, regression tests. |
More: voice-contact-center-agent (flow-spec voice agent with barge-in, verified tool calls, warm handoff, transcript eval) · ofi-cross-impact-analysis (Order Flow Imbalance cross-impact analysis).
- AI / LLM: LLM evaluation & quality benchmarking, RLHF data, AI agent workflows, RAG, prompt engineering, LLM reasoning frameworks, data curation & annotation workflow design
- Programming & data: Python, SQL, statistical analysis, predictive modeling, data visualization
- Working with clients: requirements gathering, technical writing, explaining complex concepts to mixed audiences