Field notes
Notes on backends, product development, production systems and using AI tools safely.
Building with AI and staying close to users
What a PLAYGRND WhatsApp fix taught me about faster development, understanding work at the pitch and checking whether a shipped change actually helps.
Inside the PLAYGRND video pilot: panorama, follow cam and linked goals
The engineering behind our football video pilot: Ballr capture, Core ML detections, virtual camera planning, scoreboard timing, montage and links to match events.
AI agent sandboxes: Daytona and the race to give bots a computer
A critical review of Daytona, E2B, Modal, Vercel, Fly.io, Cloudflare and AWS and the controls an agent computer still needs.
Small models in production: Shopify's flywheel and a PLAYGRND plan
How specialized model distillation works, what Shopify's 0.8B result proves, and a practical training and Mac Studio deployment plan for PLAYGRND.
x402, Cloudflare and Coinbase: what machine payments mean for Bitcoin
How x402 turns HTTP 402 into a machine payment flow, why Cloudflare and Coinbase matter, and where Bitcoin, Lightning and Grok Bot fit.
How we orchestrate 12 Codex threads on one software product
Inside the PLAYGRND workflow: twelve specialised Codex threads, one PM orchestrator, explicit ownership, compact handovers and human release gates.
How I organise overnight software delivery at Tilt
How I use Linear, GitLab, Slack and bounded local AI runs at Tilt to turn a defined software scope into review-ready changes overnight.
How to recognise an asymmetric technology bet
From the industrial revolution and early LLMs to Lightning payments: how to recognise an infrastructure shift while limiting the risk of an early technology bet.
From analogy to a testable product hypothesis
How to connect ideas across domains, detect compression error, and test the critical assumption cheaply before making a larger investment.
Block Buzz review: a Nostr workspace for humans and AI agents
A technical review of Block's Buzz, its signed-event architecture, agent identity model, practical strengths, production gaps, and fit for software teams.
From planting trees to a custom 3D planner
How a short break from prompting became a deployed site-planning tool in about 15 hours, and what that speed means for software companies.
OptMem review: a 426-token memory layer for AI agents
A technical review of Victor Taelin's OptMem, its hierarchical memory compression, practical advantages, trust limits, and fit for software teams.
What Memoato shows about trustworthy AI memory
A practical record model for keeping source evidence, reviewed facts, inferences, corrections, and agent access separate.
What an AI context audit looks like
A practical playbook for finding where AI can help a software team without giving it unsafe access, stale assumptions, or unverifiable context.
What PLAYGRND shows about building faster with AI
PLAYGRND is a practical example of how a small team can move quickly with AI-assisted engineering, as long as AI accelerates judgment instead of replacing it.
How we sped up PLAYGRND aggregates without hiding problems in cache
A PLAYGRND optimization moved stable season aggregates from raw query paths into persisted derived tables and Redis snapshot cache, with raw fallback when derived state is missing.
Why PLAYGRND uses WhatsApp for login and profile claims
For some products, the best auth flow is not another password or email loop. If users already work in WhatsApp, auth should listen to that channel.
Which context is allowed to win?
Agent reliability is not only about retrieving the right information. Software teams also need rules for conflict, freshness, ownership, and permission.
The next agent stack is accountable context
Recent agent work is converging on a simple point: memory, tools, evals, and permissions need to become one accountable context layer.
Memory is not the same as context
An agent should not merely remember text. It needs source, freshness, permission, scope, and a reason to trust that the information is still valid.