Steven M. Ledbetter

(845) 379-0641

NYC Area, NY

Summary

PM for Meta's real-time metrics and analytics warehouse in the AI & Data Infrastructure org. Scaled to absorb 9x agentic-workload traffic in 90 days while tightening latency and freshness SLAs. 20 years and three acquisitions (Supernatural to Meta, Elevate Security to Mimecast, Habitry to Lift The Bar). Co-inventor on patent US12149553B1, an ML risk-scoring and intervention system for social-engineering attacks at Fortune 500 customers.

Employment

Meta

2023 – Present

Product Manager, AI & Data Infrastructure

  • Scaled Meta's real-time metrics and analytics warehouse to absorb 9x agentic-workload traffic in 90 days while tightening latency and freshness SLAs.
  • Drove alignment and shipped a novel OLAP with watermark-based consistency SLOs; real-time freshness, <100ms latency, 100M QPS, 10x capacity savings.
  • Led Supernatural's data and dev infrastructure migration onto Meta following acquisition.

Supernatural (acquired by Meta)

2022 – 2023

VP of Product

  • Led a 17-person product organization (PMs, UXR, Program Managers) through 3x revenue and member-base growth, then through acquisition by Meta.
  • Guided technical development of real-time multiplayer infrastructure for VR fitness.

Elevate Security (acquired by Mimecast)

2020 – 2022

VP of Product

  • First product hire; built the team from scratch (PMs, POs, Designers, Technical Writers).
  • Co-inventor on patent US12149553B1, an ML risk-scoring and intervention system for social-engineering attacks.
  • Led product pivot that drove Fortune 500 adoption, 2x ARR, $8.25M Series A, and Mimecast acquisition.

Tonal

2018 – 2020

Director of Product Management, Engagement

  • Coordinated decision metrics across hardware, firmware, software, and content; 2x engagement, 80+ NPS.
  • Developed content strategy and workflows that increased production output 27x.

Habitry (acquired by Lift The Bar)

2013 – 2018

CEO, Head of Product, Co-Founder

  • Co-founded, raised $150K, shipped the first behavior-change iOS app for health coaches; sold.
Writing smledbetter.com/writing

Receipt-Gated Pipelines (2026) — Cryptographic verification of tool-call claims in multi-agent LLM systems.

Debate Degrades Reasoning (2026) — Symmetric debate degrades LLM reasoning across 2,100 evals, 11 conditions, two benchmarks.

Structure Beats Scale (2026) — Structured review outperforms brute-force generation in LLM code synthesis.

Open Source Projects github.com/smledbetter

Flowstate

https://github.com/smledbetter/flowstate

Observability platform for AI agent development. Tracks metrics, traces, gate-failure signals, drift detection.

  • Sustained 97% prompt-cache hit rate across 130 sprints in 16 projects. Only ~3% of per-sprint tokens are new work, enabling a 842M-token single-operator agent-dev pipeline.
  • Pre-registered factorial experiments; killed 2 of 3 proposed workflow changes after measuring a 10–13% noise floor.

Jig

https://github.com/smledbetter/jig

Build-time eval governance for LLM apps. No production-signal registered, no compile.

  • Closes the offline/online eval gap behind high-profile AI walkbacks (Klarna, 2025): features pass eval on green offline numbers, then regress in production.
  • Deterministic judges keep LLMs out of the audit path. Slots above existing eval frameworks (Braintrust, Promptfoo, Inspect AI) as a governance layer, not a replacement.
  • MIT, PyPI as jig-eval. 249 tests, 94% coverage. Submitted to UK AISI's Inspect AI (issue #3770).

Gas Town

https://github.com/gastownhall/gastown

Open-source multi-agent workspace manager for federated LLM task execution.

  • Identified a security vulnerability in federated completions where tool-call confabulation propagated fabricated data to downstream agents in 82% of trials.
  • Proposed and designed HMAC-signed receipt gates based on original research (Receipt-Gated Pipelines), reducing confabulation propagation to 0% across 686 trials.
Skills

Product Leadership — three acquisitions, patent co-invention, infra org transitions

Data Infrastructure — DuckDB, OLAP, real-time analytics, telemetry pipelines

AI & LLM Agents — Anthropic SDK, Claude Code, MCP, eval design, prompt caching, tool-call verification

Machine Learning — LLM-as-judge, drift detection, factorial design, noise-floor calibration

Education

Masters, Sport Psychology — John F. Kennedy University (2014)

Bachelors, Philosophy — University of Chicago (2004)