Cap
Contributor / 1 merged, 1 open
Moved the caption word utilities into a shared cap-project module so the copies that had drifted apart now agree; a follow-up makes the number of words per caption segment configurable.
PROJECT AND CREDENTIAL RECORDS
Projects & certifications
I make AI agents safe to run and test whether they actually work. Five projects show that work, followed by my open-source contributions and certifications.
01 / PROJECTS
Featured projects
Agent reliability, agent security, cloud operations, and community. Open a project for what I built, what I learned, and where the evidence is.
01 Harneloop Framework for improving agent harnesses: agents inspect what they produced, trace where it went wrong, test a candidate change in isolation, and keep it only if the test improves.
Primary outcome
Turns agent improvement into a test discipline instead of trial and error. A harness change is promoted only when a repeatable, artifact-aware test gets better, and the working baseline is never edited in place.
What I built
- Portable harness units that keep task context, artifacts, traces, evaluators, and candidate history together in an isolated environment.
- A trace-to-diagnosis-to-candidate-to-test loop that promotes or rejects each harness change on evidence.
- The evolution workflow that improved the ViperMesh Blender agent, its first published application.
What I learned
- Self-reflection is not enough when an agent cannot inspect or understand the artifact it produced.
- For narrow operational tasks, the harness around a model can be a cheaper improvement surface than fine-tuning.
- Isolating candidates stops one improvement attempt from silently breaking the working baseline.
Evidence available
- Public framework repository
- Public project site
- Technical case study
02 ViperMesh 3D workspace with a Blender agent that works through about 130 constrained, validated tools and an MCP layer instead of writing open-ended Blender Python.
Primary outcome
Against the Anthropic x Blender MCP baseline, using the same model for both, the ViperMesh harness was faster on 6 of 7 live tasks, used about 90% fewer agent tokens, and scored higher on a neutral visual rubric.
6 of 7 live tasks faster Same acting model for both harnesses.
~2.5x average speedup Across the seven comparable live tasks.
~90% fewer agent tokens Aggregate acting-agent tokens across all seven benchmarks.
What I built
- A deterministic Blender addon with about 130 task-focused tools and an MCP layer that give agents explicit scene operations and validated parameters.
- A browser 3D engine (Three.js, React Three Fiber) for inspecting generated and imported assets directly after inference.
- An open-source Blender connector and MCP server, while the wider product stays private.
What I learned
- Narrowing what an agent is allowed to do can cut token use and improve accuracy at the same time.
- Spatial tasks need inspection evidence; an agent saying it finished is not a quality gate.
- Most early failures were harness and evaluation failures, not proof the model needed retraining.
Evidence available
- Seven comparable live Blender benchmarks on the same model
- Neutral LLM-judge visual evaluation
- Public Blender connector source
03 nullstate-cli Purple-team CLI that attacks Terraform-defined infrastructure in local cloud emulators, applies a deterministic fix, and replays the attack to prove the fix holds.
Primary outcome
On AWS LocalStack, the public-bucket exploit worked before remediation and was blocked on replay after the Terraform fix. Self-hosted models explain the attack; the verdict comes from the replay, not from the model.
Blocked post-fix attack replay The exploit path succeeded before the fix and failed after it.
Private self-hosted inference Open models served with vLLM on an AMD MI300X.
48h original build window Built for the AMD x lablab.ai hackathon, continued as open source.
What I built
- A Python CLI that reads Terraform, starts sandbox targets, finds supported exposures, applies deterministic fixes, re-runs the attack, and writes evidence artifacts.
- Private model inference with vLLM on an AMD Instinct MI300X, reached through an SSH tunnel.
- A repository with PR checks, tests, type checking, CodeQL, dependency review, and a threat model.
What I learned
- A security finding is more convincing when it comes with the attack, the fix, and a replay of the same attack.
- Models can explain an attack without becoming the source of truth for the security verdict.
- Even a 48-hour hackathon build benefits from branch protection, release evidence, and stated limits.
Evidence available
- Public repository and demo video
- Live AWS LocalStack run artifacts: attack blocked after the fix
- Hackathon submission summary
Evidence still needed
- Live Azure validation, blocked by LocalStack Azure emulator access
- Complete case-study write-up
04 Private Secure AI Automation Platform & Personal Agent Private n8n platform and OpenClaw personal agent on Azure Container Apps, with private networking, managed identity, passwordless CI/CD, and remote access over Tailscale.
Primary outcome
I built and ran it, traced an idle-cost failure back to my own design choices, tore it down with full exports and inventories, and documented a cheaper rebuild. The lesson I kept: budgets and agent permissions are controls; dashboards are not.
Stateful migration Google Cloud VM to Azure containers Workflow data and configuration preserved; PostgreSQL rebuilt.
Passwordless CI/CD federated delivery GitHub Actions with Entra OIDC and managed-identity image pulls.
Scoped agent resource-group access The personal agent could operate sibling services, and nothing outside them.
What I built
- A stateful migration of n8n from a Google Cloud VM to Azure Container Apps, preserving workflows and configuration through Azure Files.
- Private networking, managed identities, and a GitHub Actions pipeline using Entra OIDC with no stored cloud secrets.
- An OpenClaw personal agent with access scoped to one resource group, sharing workflow storage with n8n so it could create and reuse automations.
What I learned
- Cost dashboards are visibility, not control. A small budget with thresholds and shutdown rules should exist before the first deployment.
- Scale-to-zero only works when the minimum is actually zero; two services were configured as always on.
- I gave an agent operational reach before the recovery boundary was ready. Destructive actions need least privilege, dry runs, and human approval.
Evidence available
- Azure cost exports by service, resource, and meter
- ARM exports and before-and-after resource inventories
- Restore scripts and a lower-cost rebuild guide
05 OpenRoot Non-profit developer community I founded: a developer program, a newsletter, and an open-source GitHub organization where members build in public.
Primary outcome
Grew from a newsletter into a developer community and program with 30+ members in its GitHub organization, a first partnership, and hackathons and workshops planned with partners.
What I built
- The OpenRoot Developer Program: a public builder guide, rewards and policy documents in six languages, and an application flow at openroot.dev with private manual review.
- The OpenRoot GitHub organization, where member and community projects are developed in the open.
- A small team: I hired and manage OpenRoot’s first social-media marketing intern.
What I learned
- A community program needs published terms, privacy rules, and a review process before it promises anything to members.
- Translated policy documents drift unless the build checks that they still match the source.
- Running a community is operations work: intake, review, partner communication, and delegation.
Evidence available
- Live site at openroot.dev
- Public GitHub organization
- Published developer-program documentation
02 / OPEN SOURCE
Open-source contributions
Fixes and features in projects I use, linked to the upstream pull requests. Each entry shows whether the work is merged, open, or was superseded upstream.
Prowler
Contributor / Merged
Fixed a 404 in the OCI audit-retention check by creating the audit client in the tenancy home region, with regression tests.
OpenWhispr
Contributor / 1 merged, 1 open
Fixed dropped keyboard taps between rows on mobile; a second PR adds automatic API-key fallback for dictation.
Excalidraw for Obsidian
Contributor and tool author / Open PR
Proposed a Focus Embed Timer script for the Obsidian Excalidraw plugin, built on focus-embed-timer, my own zero-dependency timer for Obsidian, Excalidraw, and Markdown embeds.
Graphiti
Contributor / Superseded upstream, 1 open
Traced non-OpenAI endpoint failures to client routing; equivalent routing later landed upstream. A second PR makes the MCP read tools accept scalar group IDs.
Mozilla Firefox
Contributor / In progress
Working on a hard graphics driver bug in Firefox.
Matplotlib
Mentor / Merged
Mentored a second-year computer-science student through their first upstream contribution, a merged documentation fix.
03 / MORE WORK
Earlier and supporting projects
Research, data, client, and hackathon work that shaped the featured projects. Two of them have full case studies.
PromptTriage
Prompt research and evaluation
Prompt-refinement product and research program: 1,080 structured format evaluations over a 28K+ prompt corpus, with retrieval studies and Azure ML experiments.
n8n Automation Atlas
Workflow data and retrieval
36,985 import-ready n8n workflows and a 131,648-row dataset published on Hugging Face, plus retrieval experiments that kept their negative result.
ViperMesh Azure GPU Inference Pilot
GPU inference
Hunyuan3D shape and texture services on Azure A100 and T4 Container Apps, and proof that scale-to-zero settings alone are not a GPU budget control.
Kaelux.dev
Client delivery
The business through which I delivered automation and AI/ML integration work for multiple clients, from process analysis to deployment and handover.
StoryForge and hackathon mentoring
Team delivery
Lead developer on a five-person Gemini Live hackathon team, with a 122-test backend, while mentoring an early-career developer.
Crosswind Console
Multi-agent experiment
Travel, career, and trend agents across 120+ tools; tool latency fell from about 60 seconds to under 3 with persistent MCP connections.
Kaelux Automate
Workflow generation
n8n interface for generating and editing workflow JSON with LLM assistance.
Examplayer
Learning software
Exam-practice app that turns documents into structured questions.
04 / CREDENTIALS
Certification index
Exact credential names. Microsoft SC-500 and HashiCorp Terraform Associate are next, in October and November 2026.
Microsoft Certified: Azure Administrator Associate (AZ-104)
Microsoft / Associate certification
Azure identity and governanceStorage and computeVirtual networkingMonitoring and resource administration
AI Engineering for Developers Associate
DataCamp / Associate certification
LLM application designRAGVector retrievalPrompt engineeringProduction evaluation
Machine Learning Foundations Professional Certificate
Anaconda / Professional certificate
Supervised and unsupervised learningData preparationModel evaluationPython ML tooling
Docker Foundations Professional Certificate
Docker, Inc. / Professional certificate
Container imagesMulti-stage buildsComposeVolumesContainer security and lifecycle
GitHub Foundations (GH-900)
Microsoft / Certification
Git and GitHub workflowsGitHub ActionsRepository securityCollaboration and administration
Microsoft Applied Skills: Azure Administration
Microsoft / Applied Skills credential
Hands-on Azure administrationResource configurationIdentityNetworking and monitoring
Cambridge C1 Advanced English
Cambridge English / Language qualification
Professional English communicationTechnical writingAdvanced reading and comprehension
Used across published technical work