Independent AI systems builder

Building dependable AI agent systems.

I research agent memory, build evidence-grounded workflows and Kubernetes-native runtimes, and operate a private multi-agent content system in active use.

Explore projects GitHub

Featured projects

One problem space.
Three system layers.

Each project is a public, inspectable artifact. Status labels and limitations are part of the work—not footnotes hidden after the claim.

01
Research Experimental Alpha

Memory Policy Lab

An experimental reference implementation and evaluation harness for how agents retain, compress, retrieve, revise, and assemble memory under context budgets.

DemonstratesAgent memory architecture, explicit context policy, source lineage, deterministic evaluation, and bounded provider experiments.

Current boundaryResearch software. No production-readiness, unlimited-memory, or better-than-RAG performance claim.

View repository Read limitations
02
Applied agent v0.1.0 Technical Preview

Meeting Minutes Agent

A bounded, evidence-grounded agent proof of concept that turns meeting transcripts into structured minutes with traceable source spans.

DemonstratesContext routing, exact evidence validation, conservative unknown handling, one-pass repair, unresolved isolation, and a local web preview.

Current boundaryLocal technical preview. It is not a production meeting platform or a public hosted service.

View repository Preview scope
03
AI infrastructure v0.2.0 Preview

LobsterBay

An open-source Kubernetes control plane for hosting and operating user-owned AI agent runtimes across isolated environments.

DemonstratesRuntime lifecycle management, tenant isolation, NFS/PVC-backed state, Feishu/Lark channels, Helm packaging, and observability.

Current boundarySuitable for preview and customer-site validation. Enterprise SSO and production HA remain roadmap work.

View repository Readiness notes

System view

Remember. Work. Run.

The projects form a continuous exploration of dependable agents—from internal state decisions to auditable work and real-world operation.

01

Remember

Memory policy

What should an agent retain, revise, compress, retrieve, and place into the next bounded request?

Memory Policy Lab
02

Work

Bounded workflows

How can an agent produce useful results while preserving evidence, explicit uncertainty, and human review paths?

Meeting Minutes Agent
03

Run

Runtime infrastructure

How can user-owned agent runtimes be deployed, isolated, observed, and adapted to private environments?

LobsterBay

Operating proof

Not a product demo.
A working personal system.

Hermes is included here as evidence of lived agent operations—not as an open-source release or enterprise product claim.

Private system Active since April 2026

Hermes + OpenClaw Content Pipeline

A self-hosted, single-user multi-agent workflow for daily news scanning, deep research, article writing, asset production, review, and Feishu cloud-document publishing.

Evidence snapshot · 2026-07-21
File, process, configuration, log, and output evidence.

Since Apr 2026Active personal use
78Daily topic briefs
13End-to-end pipeline runs
11Articles approved
8Feishu documents published
6MP4 files produced

What is real

Seven subagent profiles support a five-stage, filesystem-backed SOP; five have verified dispatched outputs. Ten rejection reports show review and rework in practice. Completion reports are checked against files and metadata rather than trusted at face value.

Human boundary

Topic approval, public-channel publishing, legal responsibility, and exception recovery remain with the operator. Public platforms are not automatically connected.

Current limitation

The profiles share one underlying LLM. The system is single-user, has no multi-tenant permissions or SLA, and can stall without operator intervention.

Why it matters

It demonstrates sustained use of agent orchestration, review loops, stateful file handoffs, rejection and rework—not merely a one-off multi-agent demo.

Private implementation · No public repository · Not presented as enterprise-ready or fully autonomous.

Current explorations

Beyond the public repositories.

Early-stage work remains deliberately separated from released projects. These are active directions, not product-readiness claims.

Hardening & productization

Hermes Digital Worker

Evolving the personal content pipeline toward clearer task state, timeout handling, safer tool boundaries, reusable deployment, and business-specific digital-worker patterns.

In development

Open Digital Mentor

Expert knowledge distillation and interactive digital-mentor experiences, with clean-room and asset-licensing boundaries treated as first-class design constraints.

Applied systems

Enterprise Knowledge Agents

Knowledge assistants and bounded business agents for policy search, customer service, expert support, review, and structured operations.

Working principles

Evidence before adjectives.

The aim is not to make prototypes sound mature. The aim is to make system behavior, evidence, trade-offs, and current boundaries inspectable.

01

Evidence over hype

Connect public claims to code, tests, traces, evaluations, reproducible demos, or clearly documented experiments.

02

Explicit boundaries

Alpha, preview, and research-stage work should be described as such—without borrowing credibility from future roadmaps.

03

Systems thinking

Agent behavior, context, evaluation, infrastructure, security, operations, and enterprise constraints belong in one design conversation.

About

Product judgment.
Engineering depth.
Delivery discipline.

I have spent more than a decade building software products and engineering teams, with recent work focused on LLM applications, agent systems, private model infrastructure, and enterprise AI delivery.

My work moves between research questions and delivery constraints: defining a problem, building a testable reference path, making uncertainty visible, and deciding what must be proven before a system is presented as production-ready.

Agent architecture LLM systems Context & memory RAG & knowledge Python / Go Kubernetes Private deployment Product & delivery