git log --since=2025-09-04

A year of building with AI

What I built between September 2025 and September 2026. The big ones get the full story. The rest get a picture and a paragraph, because that is what they were.

total 15 · the long builds, the tools, and the small ones

costs-portal

The md reader showing a document with its outline rail and a code block

boswell

review-kit

The Hungry Grave: a run replaying, rings of orbiting shots around the player

the-cabinet

The long builds · 5

01 / 15

boswell

drw------- · private · 463 commits · Python 3.12 · FastAPI · pgvector · Pipecat · Next.js 16

Chat assistants forget you between sessions and run on someone else's servers. Boswell keeps a continuously distilled model of one person's life, on hardware they control, and surfaces things at the moment they matter.

chat surface

> just audit-turn cf352e95
[utterance]
  user_utterance (voice): Tell me about my friend R.
[extraction]
  entity f786f5a4  person  Mark Lozano
  entity a9267990  person  R.
  edge   d7c80738  friend  f786f5a4 -> a9267990
[resolution]
  bind      person/f2b0bb7fdfdaee040e734c86187d63d5bae56e21f6c17732bd0f88e094057a7f  entity=f786f5a4 matched=f786f5a4  first_person_anchor
  bind      person/aa392d36bd86cef2f0fac485450602185be5ff071453ca350150236a18605f24  entity=a9267990 matched=a9267990  same_person_by_name
[retrieval]
  hits=8  relationships=100  entities: R., place, Mark, T., place, S., K., person
[crosscheck]
  no signal
[reply]
  [R.](boswell:turn/019f3dd1-758f-77b0-bb7b-913c263b94fd) is someone you've known since you were about 23, so for over 25 years now. You two got along really well from the start. You have a similar sense of humor, like the same foods, and both love pie. You even had a stretch where you'd celebrate "Pi Day" once a month at work, bringing in pie to share.

  You've worked together at multiple companies over the years, more than anyone else in your current friend group. He's [entertaining, nice, and generous](boswell:turn/019f3dd1-758f-77b0-bb7b-913c263b94fd). He loves hosting events and seems to enjoy making people happy.

  He has a [basement theater at his house](boswell:turn/019f3dd1-758f-77b0-bb7b-913c263b94fd) where you have movie night. He usually makes really good food for those gatherings, and the rest of you bring sides or desserts.

  You two used to game together: World of Warcraft and EverQuest before that. His wife M. also played WoW with you guys, along with [S.](boswell:turn/019f3dcd-be51-7631-962a-f66ded1d7460). R. also rides motorcycles and has introduced other friends like K. and S. through that hobby.

just audit-turn

memory browser

02 / 15

champ-sage

drwxr-xr-x · public · 424 commits · TypeScript · React 19 · RxJS · Overwolf Electron · Whisper · evalite

Every League companion app serves pre-computed tier lists that ignore your actual items, augments, and enemy team. Champ Sage is a voice-first coach for the game you are in right now.

The in-game overlay: three augment cards, each with a coaching verdict above it, and the coach answering a question in the corner

in-game overlay · augment verdicts

The desktop app in champ select: your pick with a build direction and summoner spells, the enemy team still picking

desktop window · champ select

post-game review

03 / 15

brief

drwxr-xr-x · public · 234 commits · Next.js 16 · React 19 · Vercel · WorkOS · AI SDK · [live]

Long videos hide whether they are worth watching behind the watching. Brief turns a YouTube link into a timestamped, structured summary so you can decide to watch now, later, or never, and keeps every brief in a searchable library.

A brief: the video player, the gist, and the AI-generated chapters with timestamps

a brief · chapters

The library: search, tag filters, date filters, and a grid of briefs

library · filters

> brief whoami
person@emailserver.com

> brief generate https://www.youtube.com/watch?v=vsGwx28z4jk
Fetching transcript...
Generating brief on the server... (typically 5-15s)
https://brief.niftymonkey.dev/brief/b22ee8a1-d3ad-4625-a73a-4ebfd372049f

the brief CLI

04 / 15

the-cabinet

drwxr-xr-x · public · 224 commits · TypeScript · Vite · PixiJS 8 · React 19 · Upstash Redis · [hungry-grave] · [housewarming]

My games had no shared home, and a repo per game costs an install, a tracker and a deploy each. One repo holds every game plus a launcher, and each game keeps its own stack.

The Hungry Grave

Housewarming

The replay route: a verified tape playing back at tick 9698, rings of orbiting shots around the player

#/replay · a tape at tick 9698

05 / 15

regimen

drwxr-xr-x · public · 234 commits · TypeScript · Bun · SQLite · OpenTelemetry · Grafana Cloud

Engineers judge their AI-assisted work on feel. Regimen captures what actually happened across Claude Code, Codex, Copilot, and Gemini into a local store, then tells the engineer where the work went wrong and whether their fixes helped.

> regimen rollup

Hey, so quick gut-check on the last stretch since you asked: it went well overall. Most sessions land the thing you actually asked for, roughly three in five finish clean with no real back-and-forth needed, and only a small slice, about one in fourteen, really goes sideways.

When something does need fixing, there's one pattern behind almost all of it: you catching Claude stating something as settled fact before it actually checked. A branch it called "clean" that wasn't. A claim that a skill didn't exist when it did. A "yes it's fixed" before anyone re-ran the thing. Session 6acc6193b is a clean example: Claude declared three PRs mergeable and you had to stop it and ask "are you sure?" before it actually looked. Same shape in fa647718: an audit said "no issues" while a stale line in its own doc contradicted it, and you're the one who caught the mismatch. It's not that Claude is careless across the board, it clearly does good, careful work most of the time, it's that the moment of "I'm done, here's the state" is where it gets ahead of the evidence.

My recommendation is that we add a small standing habit: before Claude reports something as verified or complete, it names how it checked, not just that it checked. That forces the "how do you know" step to happen before you have to ask it, instead of after.

The numbers behind this:
judged conversations: 425
  accomplishment: accomplished 256, not-accomplished 31, partial 63
  correction-cost: heavy 47, light 121, none 89
  verification: accepted-unverified 64, nothing-to-verify 70, verified 205

regimen rollup

> regimen status
installed: regimen 1.0.0
  codex: feedback, enforcement, guidance
  claude: feedback, enforcement, guidance
  copilot: feedback, enforcement, guidance
  gemini: feedback, enforcement, guidance
feedback: enabled
daemon: running (pid 376)
last event: 2s ago
backlog: 857193 bytes
awaiting assessment: 284 conversations

regimen status

The tools · 6

The md reader with the outline rail open beside a document with code blocks

reader · outline rail

A document in the md reader, set wide, with a Mermaid flow diagram and a code block

a doc with a diagram

md

[live]117 commits

Turn a markdown file into a legible shareable link, faster than a gist and more durable than a pastebin. The second audience is a coding agent that needs to publish and edit docs there as fluidly as it edits local files.

pickai with two capability rules applied, 854 of 1,738 models passing, ordered by rating

rule rail · results table

per-model panel

pickai

[public]93 commits

models.dev lists roughly 2,200 model identities, and most people pick the one name they have heard of. pickai takes a project's hard rules and what matters this week, then returns a short ordered list worth testing, with the reasoning attached.

a run in progress

ai-consensus

[live]79 commits

One model's answer is one opinion. This sends a question to several models at once, lets them read each other and refine over rounds, and has an evaluator model decide when they have converged.

The niftymonkey.dev homepage: a whoami line and a directory listing of projects with permission strings

ls -la ./projects/

a notebook entry

niftymonkey.dev

[live]78 commits

My personal site. I want to be easier to learn from without becoming a content creator. It lists my projects as a terminal directory listing and holds a notebook of engineering writing worth keeping.

The costs-portal dashboard for one month: total, daily average, trend, project cards with sparklines, and the watchman and long arc panels

dashboard · watchman

One project's detail page: this month's composition, six months of spend by source, and the SKU table

project detail

costs-portal

[private]64 commits

"What did this project cost me this month" used to mean checking five dashboards. This pulls Anthropic, OpenAI, OpenRouter, and the flat-fee SaaS tail into one project-first view on your own machine.

landing page

capture

review-kit

[private]51 commits

People forget what they did all year and then write a performance review from memory. ReviewKit captures accomplishments by voice or text as they happen, then turns the pile into goal-aligned review content.

The small ones · 4

41 commits

idea-vault

[live]

App ideas start as a braindump and end as scattered text. This stores them in a structure you can scan, keep updating, and export as a build-ready prompt.

refine wizard

26 commits

session-scribe

[public]

A D&D group finishes a session with a Teams transcript and a dice log and no record anyone wants to read. This turns both into a scene-by-scene narrative recap.

Session Scribe mid-generation: extraction on scene 4 of 5, elapsed time, and the activity log

generation in flight

20 commits

tool-radar

[public]

A hand-kept file of default tools went stale and got deleted. This is a catalog that refreshes itself: a weekly scheduled agent runs a written runbook and opens a PR. There is no UI. The showable artifacts are files.

The Tool Radar: 98 tools plotted on adopt, trial, assess, and hold rings across 23 areas

the radar

15 commits

tts-bake-off

[public]

Picking a text-to-speech voice by reading spec sheets does not work. This renders one line across eight engines, local and cloud, so the choice gets made by ear.

> time say.py --engine kokoro "Fifteen projects in a year. This one taught the rest of them to talk."
real    0m6.66s
> time say.py --engine kokoro "Fifteen projects in a year. This one taught the rest of them to talk."
real    0m2.52s
> time say.py --engine omnivoice "Fifteen projects in a year. This one taught the rest of them to talk."
real    0m36.00s
> time say.py --engine omnivoice "Fifteen projects in a year. This one taught the rest of them to talk."
real    0m1.68s

say.py in a terminal