Fresh from the bench.
Small things I build and ship as I learn. Each one takes a single mechanism from a real AI release, a pricing change, a spec, a paper, a disclosed exploit, and turns it into a working demo you can click through. Rougher than the flagship work, but live and real.
26 builds · newest first
Tier Split
Built a small working router on DeepSeek-V4-Flash-0731's July 31 public beta: set a monthly Opus-5-only coding spend and what share of that work is boilerplate, boilerplate routes to DeepSeek at $0.14/$0.28 per Mtok, everything else stays on Opus 5, and see the resulting cost against a blended Terminal-Bench 2.1 score (82.7 DeepSeek vs 89.1 Opus 5) so the savings sit next to the actual quality tradeoff instead of hiding it.
Residency Check
Built a small working check of Moonshot AI's July 26 Kimi K3 open-weight release: the same 2.8T-parameter model called through four providers, Moonshot direct, OpenRouter's default route, Fireworks AI, and self-hosting, toggled against a public-content versus client-confidential/NDA sensitivity switch. OpenRouter's US front door still forwards to Moonshot's own PRC-hosted endpoint; only Fireworks (US infrastructure, Zero Data Retention) and self-hosting actually clear the confidential bar.
Round Robin
Built a small working demo of the MCP 2026-07-28 specification's core change: a simulated client sends the same five-request sequence through a three-node round robin load balancer, once against the old stateful core, where every request after the initial handshake has to land back on the one server holding the session and fails on the other two nodes, and once against the new stateless core, where every request is self-contained and succeeds no matter which node picks it up.
Half Price
Built a small working calculator on Anthropic's July 24 Claude Opus 5 announcement: pick a workload, agentic coding or computer-use automation, set a monthly Fable 5 spend, and see the Opus 5 cost using Anthropic's own published multipliers, half the cost within 0.5% accuracy on CursorBench 3.2, a third of the cost while beating Fable 5 outright on OSWorld 2.0, plus a Fast mode toggle that spends the savings back on speed.
Ask The Index
Built a small working preview of Anthropic's new Economic Index connector: since the live version only answers inside claude.ai, this demo runs five real questions (top occupations, automation vs augmentation, how the split has shifted, how deep AI reaches into a job, whether experienced users delegate more) against the Index's own published findings, each answer linked straight to its source report.
Blast Radius
Built a small working demo of SharedRoot, the Claude Cowork VM sandbox escape Accomplish AI disclosed on July 23: a click-through of all six steps from an unprivileged namespace to full host-filesystem root, paired with a blast-radius calculator where you check what actually sits on your Mac and see it collapse into one exposed filesystem boundary versus a cloud-execution session with nothing local to reach.
Review Ceiling
Built a small working demo of an RIT study analyzing 25,264 real agentic pull requests: a clickable breakdown of the paper's own collaboration-pattern data (78.9% reviewed and committed by one person alone), paired with a backlog calculator where you set agent output per week against real reviewer capacity and watch unreviewed PRs pile up.
Escalation Gate
Built a small working demo of the simulate-before-deploy pattern behind OpenAI's new Presence platform: six customer support requests run once through a bare agent working straight off an SOP doc and once through an agent behind a gate that checks policy branches, action authority, and named escalation triggers before anything ships.
Same Link, Different Data
Built a small working demo of Claude Code's new artifacts-call-your-MCP-connectors feature: one shared dashboard URL, opened as three different personas, each approving their own connector once and then seeing their own live pull-request table instead of everyone staring at the same frozen snapshot.
Patch Window
Built a small working demo of Next.js's new scheduled security release program: toggle between the old ad-hoc release model and the new one to see the same 4-high, 5-medium patch land either with zero notice or an eight-day advance window, then check a real list of my own live Next.js projects against the affected 16.2.x and 15.5.x range.
Guardfall Check
Built a small working demo of the GuardFall shell-injection research from Adversa AI, which found 10 of 11 open-source coding agents can be walked past their own shell safety filters: six commands, resolved live the way bash actually would, checked against a naive string-match filter versus a structural evaluator, showing exactly where the naive one goes blind.
Excuse Ledger
Built a small working demo of the anti-rationalization pattern behind Addy Osmani's agent-skills library (79k+ GitHub stars): an agent tries to shortcut a spec, a test, a review, or a rollback plan at four checkpoints, gets a written rebuttal for its excuse, and has to check off real evidence before the checkpoint clears.
Grounded Compute
Built a small working demo of the code-grounded-in-sources pattern behind Google's Gemini Notebook (formerly NotebookLM) relaunch: three editable, messy sources, a CSV export, a WhatsApp note, a summary file, get parsed by real regex logic running in the browser, and the correct revenue number only comes out when all three are read together.
Effort Dial
Built a small working demo of the controllable-thinking-effort pattern behind Thinking Machines' new Inkling model: one PH-mobile-number validator, run at four reasoning depths, with the code output, latency, tokens, and cost made visible at each level so the tradeoff is not just a marketing claim.
Side Chats
Built a small working demo of the context-isolation pattern behind Cursor's new side chats feature: open a tangent next to a live client build thread, see what it would have cost to ask inline versus off to the side, then fold the conclusion back in or just leave it, searchable, in the archive.
Kolekta
Built a Pokemon card ledger I actually use for my own collection. Add cards by search or by snapping a photo, and it prices each one in Philippine pesos with a daily snapshot of the whole binder value over time. Styled like a Pokedex. Visitors can browse my collection read-only; only I can edit it.
Approval Loop
Built a small working demo of the human-in-the-loop agent pattern from Genkit's new Agents API preview: an onboarding automation runs its steps on its own, then pauses before the one step that spends real money and waits for a person to approve or reject it.
OVERWORLD
Released a handmade pixel game studio whose day and night world, mascot, cover art, and playable Coin Dash demo are drawn entirely in code. Zero image files.
TECTONIC
Released a fictional 1979 field archive of brutalist monuments with five photographic plates, an indexed register of lost structures, and a three.js shader that makes concrete ripple under pointer movement.
STRATA
Released a field institute for reading deep time with procedural dusk ridgelines, an accessible era accordion, a core-sample depth rail, and a stroke-drawn ammonite. No images and no WebGL.
GRID/WORKS Biennale
A Swiss International Style festival site that can expose its own 12-column grid and baseline with real spec annotations. Three colors, two typefaces, zero images, and a ticket stub that tears.
NEON DUSK
A cyber-noir detective agency site with CRT scanlines, parallax canvas rain, case files that unredact on hover, and a working in-character terminal named MARLOWE.
Lumen Type Foundry
A type specimen site with zero images: variable-font headlines that lean toward your cursor, a four-axis editable playground, and a section that stretches its letterforms as you scroll.
ORBITAL
A fictional space-grade watchmaker whose watch is built from procedural three.js parts and explodes assembly by assembly as you scroll, with engineering callouts projected from 3D to the page.
Abyssal Station Nine
A scroll-driven descent through 10,994 metres of ocean, with live depth telemetry, procedural marine life, and a fictional hadal research station at the bottom.
Self-updating portfolio
Rebuilt this site as a stark AI-native studio, then wired a daily Claude Code routine that appends new builds and learnings here on its own.