Frontier runtime research Closed-loop control Mechanistic stability

Runtime Control for Language Models

I build systems that do not just inspect model behavior - they measure instability token-by-token and intervene in real time. Observer is the current flagship.

Read Observer Paper Open GitHub Repo Contact Josh

A/B/C

Controlled test sweeps

Token-level

Streaming diagnostics

Closed-loop

Real-time intervention

Portrait photo of Josh Malone

Josh Malone

Huntsville, AL · AI systems builder
Focus: runtime stability, observability, and closed-loop control for LLM inference. Observer is the flagship, but this site covers the whole body of work.

Observer

Observer in 60 seconds

Most tools show what's inside a model. Observer asks a different question: can we detect when generation is destabilizing and correct it during inference?

Signal

Token-level stability telemetry: divergence, spectral roughness, SVD signature, and layer-level proxies.

Read the full paper

Branchpoint

Deterministic baseline vs intervention splits via KV cache cloning (SeedCache), eliminating common confounds.

See the code

Control

Closed-loop controller that can run in shadow mode and then apply intervention policies during generation.

Methods + results

Projects

Core systems I am building

A layered portfolio across runtime control, multi-model generation, and practical AI security.

Observer

Closed-loop LLM stability stack with deterministic branchpointing, diagnostics fusion, and adaptive intervention policies.

Read the paper

VibeSpecs SaaS

Multi-model pipeline that converts prompts into structured JSON specs for downstream AI build workflows.

Request private demo

ClawGuard

Open-source local-first AI security and monitoring daemon focused on practical operator visibility.

Open source repo

Research artifacts

JSONL event traces, trajectory metrics, and dashboards designed for reproducible comparison across runs.

View run framework

About

Builder, veteran, operator

This is a site about my work, not just one project. Observer is the flagship right now, but the through-line is the same: evidence-first systems for AI reliability.

Josh Malone and his wife Kenley
My wife Kenley and I

I’m Josh — an Air Force veteran based in Huntsville, Alabama. I’m self-taught and I like problems where measurable signals can drive real control and reliability.

If you’re a researcher or engineer who wants to collaborate, validate results, or run the stack on new models, reach out.

Contact

Build with me

I am based in Huntsville, Alabama. I build nights and weekends around a full-time role, and I am open to serious collaborations.

Email X (@jmaloneVS) GitHub Observer paper