Peleke Sengstacke

I build agent systems that learn from production usage and the observability to prove they're working and improving over time.

Can agents learn to optimize their own context?

Agents pick what goes into their own context window. When they pick well, they get better. When they pick badly, they waste your money.

This research program builds the infrastructure to measure which is happening and push it toward the former. Open-source and operational.

View the full research program →

Independent Research & Engineering

October 2024 – Present

  • Building Vindler/qlawbox: instrumented agent runtime with Bayesian learning, causal knowledge graph, and post-deployment observability
  • Overclock Accelerator (Sep 2025 – Jan 2026): SME for practitioner-facing agentic engineering curriculum
  • Shipped Interlinear (adaptive language learning) and LinWheel (agentic content pipeline) to stress-test the stack on real users; built Swae OS (federated health platform, alpha) as a deeper architecture and CI/CD proof

Principal Engineer

$60MM+ ARR · 60+ NPS

edX / 2U

March 2021 – September 2024

  • Architected distributed sandbox platform for Cybersecurity Bootcamp serving 40+ universities internationally
  • Built golden VM images (Ansible + Packer) with CI-managed rebuilds into Azure VM registries
  • 10x provisioning performance via async-first migration; 66% cloud compute cost reduction through lifecycle management

Lead Architect, Cybersecurity

acq. $750M

Trilogy / 2U

March 2019 – March 2021

  • Created dedicated engineering function; hired, trained, and managed 3 engineers
  • Shipped sandbox platform serving 40+ universities; revenue grew to $60MM+ ARR
  • Early experiments in Dockerized lab environments for secure, distributable sandboxed execution

Lead Curriculum Architect

Trilogy / 2U

October 2018 – March 2019

  • Owned, built, and shipped 24-week Cybersecurity Bootcamp (Trilogy's 4th product line)
  • Co-instructed first cohort at Georgia Tech

Lead Curriculum Engineer

Trilogy

July 2016 – October 2018

  • First full-time curriculum engineer hired to build content; ended up building programs
  • Authored coursework: Django, Laravel, MERN, testing, data visualization, ML
  • Top-rated instructor at Rutgers (Full Stack) and Georgia Tech (Cybersecurity)

Interlinear

Adaptive learning platform

Course generation from arbitrary text sequences with morphosyntactic error classification feeding Thompson Sampling. CLTK + Reynir NLP pipelines for Classical Latin and Icelandic.

CLTK · Reynir · Mastra · Next.js 15 · Supabase

Swae OS

Federated multi-service health platform

Seven FastAPI services behind a Hive Gateway federated supergraph, shipped to Cloud Run by a 12-workflow workflow_run CI/CD chain (OpenTofu plan/apply, Workload Identity Federation, no stored keys). Functional alpha.

Expo · GraphQL Federation · Hive Gateway · OpenTofu · FastAPI · Cloud Run

LinWheel

Agentic AI content surface

Vindler extension with 17 agentic tools for LinkedIn content pipeline. Voice profiles, approval gates, and a learning loop that improves from feedback. 933 tests passing.

Vindler (OpenClaw) · bilrost (STRIDE sandbox) · Claude · FAL Flux.1

Princeton University

2012–2014

AWS Certified Solutions Architect – Associate

2023–2026

← About