Justin Bronder
Applied AI & Forward-Deployed Engineering
30+ years shipping production ML; wrote the FDE playbook at Microsoft as the org scaled from 10 engineers to 1,000.
Featured Research
Instrument Effects in Language-Model Honesty Evaluation
Evaluation design choices can manufacture the verdicts we attribute to models. The paper demonstrates this with an auditable single-system harness and proposes a four-check integrity protocol for honesty evals.
Work
Evaluation Integrity
Research on how evaluation instruments shape their own results, published as the instrument-effects paper above. The companion substrate, InspectAI, builds on UK AISI's Inspect framework to instrument verification degradation in long-context sessions, where checks stop firing as rapport accumulates.
Agent Memory
An MCP server providing persistent, cross-context memory for Claude, in daily production use as working infrastructure rather than a demo. Vector retrieval sits alongside a provenance store with human-gated deduplication and keyed exact recall.
Private repository; demonstration available.
Production Agentic Systems
A multi-tenant platform on Cloudflare Workers, including a WhatsApp agent, live with real users and per-tenant deployment configuration. Built and operated end to end.
Selected History
- 2014–2020 Microsoft forward-deployed engineering, founding team. Customer engagements included Delta Air Lines, 20th Century Fox, the United Nations, and nation-scale mine detection in Ghana.
- 2008–2013 Cardinal Optimization. Primary implementer of all-streets routing algorithms used by Navteq, Google, and Microsoft.
- 1990–1995 PhD work in computer science at UC Davis, resident at Lawrence Livermore National Laboratory.
- 1987–1988 Early neural network work: an undergraduate paper on neural networks, then a Hopfield network for handwritten digit recognition, a decade before MNIST was public.
Writing
The Trajectory Gap (LinkedIn, April 2026). Argues that enterprise AI deployment is bottlenecked not by model capability but by the coordination layer between long-horizon human intent and model context.