Writing / 19

Build to learn
write to remember

Notes from projects I built myself: framing the question, choosing tools, working with data, and checking the result.

Choose a
way to read

Start with the open project or go straight to the question that interests you.

01 / One build in public

An agent, SQL, and a bounded warehouse

The current article moves from routing to evaluation, with a repository and walkthrough to inspect.

Read ↗02 / Browse by question

Start with the problem you have

Move into agents, data, evaluation, or the economics of running a query.

All writing

Browse by publication date

Field note · Aug 01, 2026P0 / Research

Agent density is not operational efficiency

Running more agents on one VM matters only when failure rate, queue time, and cost per completed task are measured with it.

Agent infrastructure4 min read

Field note · Aug 01, 2026P0 / Research

AI repositories need two gates: permissions and licensing

A security workflow is not automatically a complete control, and a model repository without clear licensing remains a product risk.

Open-source risk4 min read

Field note · Aug 01, 2026P0 / Research

Cash-flow data is becoming a credit file

As SME underwriting moves beyond collateral alone, clean operating data becomes a financing and negotiation capability.

SME finance4 min read

Field note · Aug 01, 2026P0 / Research

Grok Build and Colibri belong in an experiment queue

Rapid star growth is a discovery signal; pinned commits, tests, issue activity, and rollback determine whether a repository belongs in the stack.

GitHub trending4 min read

Field note · Aug 01, 2026P0 / Research

Landed cost is not the freight quote

The cheapest quoted route can cost more after handling, inventory, lead time, surcharges, and customs obligations are included.

Landed cost4 min read

Field note · Aug 01, 2026P0 / Research

Model routing should start with cost per task

Token prices become useful only when they are connected to the quality, latency, and risk required at each workflow step.

Model economics4 min read

Field note · Aug 01, 2026P0 / Research

Open weights are not production readiness

Long context and large parameter counts create a test queue; licensing, hardware, latency, and Vietnamese quality determine real fit.

Open-weight models4 min read

Field note · Aug 01, 2026P0 / Research

Rising labor cost does not automatically mean cuts

Before reducing headcount, compare training, automation, outsourcing, and work redesign by cost per accepted unit of output.

Workforce economics4 min read

Field note · Aug 01, 2026P0 / Research

Sales agents cannot fix fragmented CRM data

Agents can accelerate research and follow-up, but duplicates, missing fields, and inconsistent pipeline definitions will amplify error.

AI agents4 min read

Field note · Aug 01, 2026P0 / Research

Skills-based hiring starts with the workflow

A stable labor market does not make hiring easy; the workflow must be translated into observable skills and work samples.

Skills-based hiring4 min read

Field note · Aug 01, 2026P0 / Research

Supply-chain resilience is a management decision

Oil shocks, route disruption, and weaker demand need owners, switching thresholds, and priced alternatives, not only logistics alerts.

Supply-chain resilience4 min read

Field note · Aug 01, 2026P0 / Research

Tax deferral is not tax saving

More time to pay can relieve cash flow, but it does not erase the underlying liability or create guaranteed savings.

Tax timing4 min read

Field note · Jul 30, 2026P0 / Research

Autonomous agents need boundaries outside the prompt

A research note on why tool permissions, network egress, credentials, checkpoints, and human review must carry the safety load.

AI agents5 min read

Field note · Jul 30, 2026P0 / Research

AI changes work most when the workflow changes

Why adding a chatbot to old handoffs is weaker than redesigning decisions, roles, escalation paths, and measures around human-AI collaboration.

Future of work5 min read

Field note · Jul 30, 2026P0 / Research

A model score is not an infrastructure decision

Why model and hardware leaderboards need workload traces, configuration context, and a local task benchmark before they can guide procurement.

AI infrastructure5 min read

Field note · Jul 30, 2026P0 / Research

Finance AI should prepare decisions, not release payments

A control-first pattern for using AI in invoice exceptions, approvals, reconciliations, and financial reporting.

Finance operations5 min read

Field note · Jul 30, 2026P0 / Research

Forecast accuracy is not the inventory decision

Why safety stock, service level, lead time, holding cost, and exception handling must sit beside the forecast metric.

Supply chain5 min read

Field note · Jul 30, 2026P0 / Research

Trending is discovery, not open-source due diligence

A practical evidence gate for adopting a fast-rising repository when AI agents and CI workflows can amplify supply-chain risk.

Open source5 min read

Field note · Jul 20, 2026P2 / Direct use

Building an Olist Ecommerce Analytics Agent with Google ADK and agents-cli

A firsthand look at an Olist and BigQuery multi-agent system: specialist routing, query guardrails, and what two local evaluation suites did and did not prove.

AI Agents7 min read