harness-factory.md

I build and own my agent harness: the prompts, tools, context routing, and review loops that turn an LLM into a system that does real work, unattended.

Technologies

• pi, a harness I can change all the way down
• Custom skills, roles, and tools
• Subagents and fresh-context review loops
• Prompt templates and context routing

Proof

The whole system runs: /factory.

disposable-orchestrators.md

I architect agent work so state lives in the substrate, not the agent. Each orchestrator reads a graph of goals, makes one move, writes its decision back, and dies, so any agent can crash and the work picks up where it left off.

Technologies

• Work as a dependency graph, not a chat log
• Tiered memory kept by lifecycle
• Disposable orchestrators over shared substrate
• Crash-resumable, parallel-safe workflows

Proof

The work loop behind /factory, including a planner that proposes the next move.

sandboxed-autonomy.md

I run agents unattended by putting hard limits around them: containerized execution, every shell command checked before it runs, and credentials the agent never holds directly.

Technologies

• rip-cage, a container plus command-gating I built
• Credential isolation through a proxy
• Always-on host, not a laptop
• Safe-to-run-unattended execution

Proof

Open source: rip-cage.

llm-app-engineering.md

Beyond tooling my own workflow, I build agentic apps: LLM calls wrapped in tools, structured output, and guardrails meant to hold up in production.

Technologies

• LLM-powered features and agents
• Tool use and structured output
• Retrieval and context design
• Evaluation and guardrails

full-stack-web.md

I build full-stack web apps across backend, frontend, and the data layer behind them. I default to hypermedia-driven apps: server-rendered HTML with focused interactivity and little client-side state. I build with React too, but reach for it when an app genuinely needs a rich client.

Technologies

• Django, PostgreSQL, MongoDB
• HTMX, Tailwind, Astro
• React, JavaScript

Approach

Why hypermedia-driven: htmx.org/essays/hypermedia-driven-applications.

apis-and-deployment.md

I build APIs, containerize apps, and deploy them with practical CI/CD.

Technologies

• Django REST Framework
• Pydantic
• Docker
• DigitalOcean
• GitHub Actions

analytics-instrumentation.md

I instrument product and marketing data so teams can trust what they measure.

Technologies

• PostHog
• GA4
• BigQuery
• Dataform
• Looker Studio

bayesian-modeling.md

I use Bayesian modelling when signals are noisy and decisions still need to be made.

Technologies

• Python
• PyMC
• pandas
• NumPy
• Marketing Mix Models

Proof

Built a Bayesian MMM at Moonbird.

computer-vision.md

I worked on medical image processing, 3D meshes, and computer vision systems.

Technologies

• C++
• Computer vision
• Medical image processing
• 3D mesh processing
• Deep learning consulting

Proof

Created MeshMonk; used or cited in 140+ research projects.

data-pipelines.md

I build ELT pipelines that turn source data into models, dashboards, and decisions.

Technologies

• Python
• SQL
• BigQuery
• Dataform
• API and database extraction

Proof

Built the Moonbird pipeline behind growth reporting and modelling.

performance-marketing.md

I combine acquisition, experiments, automation, and measurement into one growth loop.

Technologies

• Paid search
• Paid social
• Email marketing
• SEO and CRO
• Marketing automation

Proof

At Moonbird, grew monthly revenue from four to six figures in 3.5 years.

server-side-tracking.md

I preserve attribution signal by sending conversion events from the backend.

Technologies

• Server-side tracking
• Conversion APIs
• Backend event forwarding
• Ad-platform measurement

Proof

Built and maintained server-side tracking for Moonbird's ad stack.

bayesian-mmm.md

I model channel contribution when last-click attribution is too noisy to trust.

Technologies

• Python
• PyMC
• Bayesian modelling
• Marketing Mix Modelling
• Diminishing-return and lag effects

Proof

Used at Moonbird to guide marketing budget allocation.

strategy.md

I founded and ran an impact-first organization focused on digital jobs and training.

Methods

• Vision and strategy
• Market selection
• Partnership design
• Funding approach

Proof

Humainly supported 30 people through NGO partnerships in Venezuela and Nigeria.

teams-and-ops.md

I ran small international teams with explicit roles and distributed authority.

Methods

• Holacracy
• Hiring and onboarding
• Role design
• Remote operations

Proof

Humainly operated across Belgium, Venezuela, and Nigeria.

learning-from-failures.md

Humainly did not scale before founder burnout; that changed how I think about pace and sustainability.

Methods

• Retrospective practice
• Scope control
• Founder sustainability
• Impact-model evaluation

Proof

The mission worked; the growth rate and operating model did not.