Paul Santillan

Applied AI Engineer

Agentic Platforms · Production LLM Systems

Summary

Co-founder and engineer of AdBliss, a multi-tenant AI-agent platform — an autonomous agent, an MCP server, and seventeen OAuth integrations, pre-revenue and in private beta with four design partners. Previously the only data hire at a DTC subscription company, where I shipped a production OpenAI support agent in July 2023 that has since handled 140,000+ conversations at roughly 70% resolution without human escalation. Self-taught, with seven years of production data and systems work at Uber, Lyft, and startups before that.

Experience

Co-Founder, Sole Engineer, AdBliss · Remote

2025 – Present
  • Solo-built a multi-tenant AI-agent SaaS end to end on Cloudflare's edge stack — seven Workers across a 13-package monorepo, a 160-route REST API, and a React frontend — pre-revenue, in private beta with four design partners.
  • Implemented the agent as a two-persona ReAct loop on durable, checkpointed Cloudflare Workflows, with a validation layer that checks every proposed action against database ground truth before it can reach a production write.
  • Shipped a remote MCP server exposing a 27-tool registry with an OAuth consent flow, plus PKCE connectors for seventeen ad and commerce platforms with field-level credential encryption.

Senior Data Scientist, Unagi Scooters · Remote

Jan 2023 – Present
  • Sole data hire, on the Finance team; part-time since 2025. Shipped an OpenAI-based support agent in July 2023, embedded in the Shopify/Stripe subscription portal — 140,000+ conversations over three years, roughly 70% resolved without human escalation, still in production.
  • Built a credit-risk check on LexisNexis data with a risk-scaled deposit policy; subscription delinquencies fell roughly 90% without turning away the largely subprime customer base.
  • Trained a PyTorch DeepSurv churn model and built the data stack end to end (Snowflake, Fivetran, Sigma).

Operations Coordinator — Data & Analytics, Lyft · San Diego

Nov 2020 – Nov 2022
  • Owned KPI reporting and deployment strategy (Trino SQL, Python); the demand-based deployment system produced the #2 ride volume in San Diego on roughly a quarter of the market leader's fleet.
  • Reduced theft from roughly 300 to 15 units/week through hotspot detection, high-risk corral flagging, and live rebalancing; impound losses fell to near zero.
  • Traced stolen units to an out-of-state gray-market rental network using Bluetooth-tracker data and photo ID matching, supporting a Lyft security / Texas DPS operation that recovered 1,500+ scooters.

Global Security Data Specialist, Uber (Jump) · San Diego County

Feb 2019 – Sep 2020
  • Promoted twice in roughly three months; final role covered theft and loss analytics across seven-plus markets. Found a core ride-count KPI overstated by roughly 10× — a SQL filter with no market predicate was attributing every city's scooter rides to San Diego — leading to the analysis behind that market's shutdown.
  • Found and reproduced a hardware exploit enabling mass scooter theft, then worked the rollout of the tamper-resistant fix; losses fell from roughly 5,000 to 1,200 units per comparable period.
  • Reconstructed a theft ring's route from correlated GPS dropouts, contributing to a $150,000+ recovery, and presented the GPS reconstruction in court.

Skills

LLM & Agents. Agent architectures (ReAct, multi-persona), tool calling, Model Context Protocol, prompt engineering, output validation & guardrails, context compression, eval design, OpenAI API.

Languages. Python (pandas, NumPy, PyTorch, lifelines, scikit-learn), TypeScript / Node.js, SQL (Trino/Presto, Snowflake).

ML & Data. Survival analysis (DeepSurv, Kaplan-Meier, Cox PH), churn and revenue modeling, credit-risk modeling, anomaly and hotspot detection.

Platform. Cloudflare Workers, Durable Objects, Workflows, Queues, D1; Snowflake, Fivetran; OAuth 2.0 / PKCE; GitHub Actions, Playwright, React.