JD Davenport
I founded and run the BYU AI Foundry, and I’m an MBA candidate at BYU Marriott. I’ve spent 8+ years in product, engineering and consulting at National Grid, Siemens and Deloitte. On the side, I build Nerve Center, my own system of 30+ AI agents.
Dad of 2 in Provo, Utah
Open to AI Product Manager roles. Available to start now.
- $2M to $20M+internal platform chargeback, annualizedNational Grid, 2022 to 2025
- 8client engagements, from local startups to BCG XBYU AI Foundry, now
- 3% to 65%Spark AI active adoption in 10 weeksSiemens, summer 2026
- 30+agents in daily production, 100+ built over timeNerve Center, my own agent system
How I run production AI agents.
Nerve Center is my own multi-agent system: 30+ agents in daily production, 100+ built over time. The drawing shows structure only. Every part fails eventually, so every failure needs a catch; the lettered marks are four of them.
Read the drawing as text
- The operator (me, on phone or desktop) works through the cockpit, a TypeScript UI.
- The cockpit puts work on a message bus, which queues it. A frozen modal here is caught by the watchdog (C).
- A supervising daemon runs the agent sessions and restarts any that die (A).
- Inside the sessions, one CEO agent fans work out to 30+ specialists in daily production.
- Sessions retrieve from a memory graph first, and read and write durable state, where tombstones reject a stale tab’s write (B).
- Anything that speaks for me to another person (email, messages, posts) waits at one send gate for my yes (D).
- A watchdog layer wraps everything and measures work done, not uptime.
A: Sessions die. A guard brings them back.
Agent sessions crash and get killed mid-task. The supervising daemon notices the missing session and restarts it from durable state, so the work resumes instead of quietly vanishing.
B: A stale tab cannot overwrite the present.
A closed session leaves a tombstone. When an old browser tab tries to write to it later, the write is rejected rather than bringing a dead session back to life.
C: A frozen modal cannot stall the queue for long.
One blocking dialog can freeze a session and everything queued behind it. The watchdog measures work completed, not uptime, so it sees the stall, clears it, and the queue drains again.
D: Nothing speaks for me without passing the gate.
Agents draft email, messages and posts. Anything that speaks for me to another person waits at one gate for my yes. None of it skips the gate.
The incidents behind these guards, and what changed after each one, are written up in the agent safety case study .
Things I’ve shipped in the open.
Public repos you can clone. The first row is one arc: the case study, the core, and the same controls packaged for an enterprise.
-
Case study · start here
agent-safety-case-study
Each capability, its risk, the control that bounds it, and one incident where the control failed.
-
Open source · MIT · Python
orchestra-agents
The production core as installable code: bounded fan-out, a durable message bus that dead-letters loudly, circuit breakers, best-of-N with an adversarial judge, and permission tiers from read to irreversible.
-
Open source · MIT · Python
claude-deploy-kit
The same controls shaped for an enterprise: one deny-by-default policy chokepoint, a destructive-command hook that holds under skip-permissions, a fail-closed eval gate, and a redacting audit log.
Also shipped
- mcp-judgeA calibrated multi-persona LLM judge as an MCP server, shipped with the eval harness that proves the calibration. Open-sourced through the BYU AI Foundry.GitHub
- claude-bug-squashTriages a repo, fixes in an isolated worktree and proves red to green, behind blast-radius caps, a deny-list and a three-seat review. It can never merge.GitHub
- deep-research-agentIterative research with a knowledge-gap loop, nine parallel sources, and per-claim refute-or-survive checks.GitHub
- Nerve CenterThe private system all of this came out of. It runs my real work, so it stays closed; the teardown and a capabilities page it regenerates daily are open.TeardownCapabilities
More, including the engine behind the private system, at github.com/jddavenportOpen .
Start here, by how much time you have.
-
10 seconds
The short version, ready to forward
JD Davenport is an AI product leader with 8+ years across product, engineering and consulting: a machine learning consultant at Deloitte shipping production ML for Pfizer and GM, Principal Product Manager at National Grid, and now Founder & Managing Director of the BYU AI Foundry, with 8 client engagements won so far.
At Siemens in summer 2026 he took Spark AI from 3% to 65% adoption in 10 weeks; at National Grid he grew an internal cloud services platform from $2M to $20M+ in annual chargeback.
He runs Nerve Center, his own system with 30+ agents in daily production. jddavenport.com
-
5 minutes
The Nerve Center teardown
One system taken apart: what I built, how it fails, and what I decided along the way.
-
Go deep
The docs
Plain-language docs on core AI skills, patterns and the architecture behind the drawing above.
Four doors, by what you came to do.
-
Read
Raising Clawd, my newsletter on building agent systems that hold up, and the full production teardown.
-
Learn
Plain-language docs on core AI skills, in order. Students who want to ship real work can apply to the Foundry.
-
Build with me
Foundry client work: a 30-builder founding cohort that has shipped 4 production AI products and won 8 client engagements. Also speaking.
-
Press
JD Davenport is Founder and Managing Director of the BYU AI Foundry, where a founding cohort of 30 student builders has shipped 4 production AI products and won 8 client engagements. He is a BYU Marriott MBA candidate (class of 2027). Before the MBA he was a Principal Product Manager at National Grid (2022 to 2025), where he grew an internal cloud services platform from $2M to $20M+ in annual chargeback and brought a generative AI platform to 3,000+ monthly users. In summer 2026, as an AI Product Strategist at Siemens, he took Spark AI from 3% to 65% adoption in 10 weeks. Earlier, as a machine learning consultant at Deloitte, he shipped production ML for Pfizer and GM. That is 8+ years across product, engineering and consulting. He also builds and runs Nerve Center, his own system of 30+ AI agents in daily production.