AI Hijinx · The Ledger · 2026

One human, two laptops, and a payroll of agents.

We tend to picture AI work as one chat window and a lot of copy-paste. Mine is a monorepo with 117 project folders, 53 live git worktrees, and 130 custom skills, run by agents that keep 238 memories about how I like work done. Most nights the loudest thing in the room is a gaming laptop’s fan spinning up, because a free local model is narrating a video in Bangla (rent: $0). This page is the ledger of all of it: what got built, what it did in the real world, and what it cost.

Open the ledger Back to the main site

The numbers

Measured off the machines, not remembered.

Every figure below came from counting the real thing on 2026-08-15: the directory tree, the git logs, the spend ledger, the memory store. When the agents are asked what something cost, they quote the ledger. So will this page.

0Custom skillsone-command procedures the agents run, from deploys to demo films

0Project foldersone monorepo, every venture and side quest inside it

0Live git worktreesparallel feature branches checked out right now

0Commits in 2026across the five main repos, Jan through mid-August

0Memories on filehard-won facts the agents keep so mistakes get made once

0Media assets generated988 images + 1,119 voice clips, every one logged with its cost

$0.00Total media spendfor all 2,107 of them (most run free on a gaming laptop's GPU)

0Demo films publishednarrated product walkthroughs, recorded by the test suite itself

The receipts

Where the work actually landed.

Three tallies straight from the logs. The venture repo dwarfs everything, which is how it should be. The media factory's bill fits inside a lunch.

Commits in 2026, by repo

git rev-list --count, Jan 1 to Aug 15. All branches.

olive-analytics2,231

the main Olive venture repo

khwajalabs-v2711

the learning platform rebuild

khwajalabs (legacy)403

the original platform

olive-sonic219

the in-house voice engine

ops monorepo158

shared skills, deploys, portals

The media factory's output

Every generated asset appends a line to a spend ledger. 2,107 lines so far, $43.51 all-in, because the heavy lifting runs on a local GPU.

Voice clips (TTS)1,119

Images988

53 worktrees, right now

Parallel feature checkouts alive on disk as this page shipped. Each one holds a handoff file so any fresh agent can resume it.

olive-analytics35 checkouts

olive-sonic14 checkouts

everything else4 checkouts

The grid

Two laptops and a cloud, on speaking terms.

The Mac holds the truth. The gaming laptop across the room does the free GPU work. AWS holds whatever needs to face the public. A private mesh network keeps the three in sync, and a sync skill settles any disagreement in the Mac's favor.

The Mac

Source of truth

130 skills and 238 memories live here
Every agent session starts here
Pushes skills, docs, and memory outward

tailscale

The Victus

Free GPU labor

Local image generation, no per-image bill
Bangla TTS and heavy model runs
Its conversations sync back for the record

deploys

AWS

The public face

The ventures' apps on ECS
Demo portals and this very site on S3
Live phone numbers wired to voice agents

117 project folders53 live worktrees5 git repos in daily rotation1 spend ledger0 lost handoffs (the context DB insists)

Ledger entry 01

Olive: voice AI, down to the SIP trunk

Olive's voice AI ran on a vendor charging 13 to 31 cents a minute. So we built our own engine, from the SIP trunk up, and made it speak the vendor's exact API so switching over is two config values and switching back is deleting them.

on disk:olive-sonicolive-bangla-ttsolive-voice-outreach+ 14 worktrees

A voice engine from the trunk up

Real telephone calls arrive over SIP trunking (Telnyx), land in a LiveKit media room, and get answered by a speech-to-speech model on AWS Bedrock, with an OpenAI fallback if it stumbles. Number purchasing, porting, caller-ID rules, voicemail detection, mid-call agent handoffs, live call control: all in-house.

receipt: First live call measured at ~3-4¢/min against the vendor's 13-31¢. The model's median response time: 109 milliseconds.

SIP trunkingSpeech-to-speechWire-compatible

Agents that interview each other

Before an agent update ships, nine persona agents call it: the terse owner, the hostile one, the Spanish speaker, the code-switcher, the gatekeeper, the phone-menu maze, the voicemail. Hard checks (did it take turns, did it go silent, did it end naturally) gate the deploy. Timed barge-ins test whether it recovers from being interrupted mid-sentence.

receipt: A full 6-scenario release suite runs for 22 cents. Its smoke test caught, in 6 seconds, the exact regression that had once taken down live calling for a day.

Agent-vs-agentDeploy gateMeasured

Two engines in a measured bake-off

A second engine (speech-to-text, then a text brain, then cloned-voice synthesis) runs head to head against the flagship speech-to-speech model. Same test suite, same personas, per-agent switchable. The dashboard charts hearing loss, interruption recovery, language-switching, and cost per engine.

receipt: The cascade now equals or beats the flagship on every measured axis except reflexes (1.8s vs 136ms median) at roughly half the model cost per call.

A/B by instrumentVoice cloning

Building a Bangla voice

No commercial TTS ships good Bengali, so we're fine-tuning our own: 275 hours of permissively-licensed Bangla speech catalogued (with a provenance record), a text normalizer that reads Bangla numbers, currency, and OTP digits aloud correctly, a four-model bake-off rig, and LoRA pilot fine-tunes on the laptop GPU before renting big iron for production runs.

receipt: The data-prep pipeline ran on the same 8GB laptop GPU that was simultaneously rendering video frames at 97% load. Both finished.

Fine-tuningBengaliOwn the stack

The outbound campaign, run honestly

AI agents cold-call real Chicago auto shops about real cars. The rules are strict: no voicemails, every phone number web-verified first, and no call ever labeled 'interested' unless the transcript quotes the human saying so. The outcome classifier was rewritten to demand evidence. Shops that said no once are excluded forever.

receipt: A 99-call campaign day ran unattended from the gaming laptop. The 66 calls where our agent (not the shop) fumbled sit in a redial ledger, on the record.

OutboundTranscript-audited

The prospector and the demo machine

One agent mapped 3,347 Chicago SMBs into a tiered CRM. Another builds prospect-customized demos: research the company, clone an interactive workspace with their facilities and catalog, film a personalized narrated walkthrough, publish it behind auth.

receipt: 10 demo films live on the demo portal. The live demo line answers right now: 855-614-0006.

3,347 SMBsLive line

Ledger entry 02

KhwajaLabs: proof of skill, not homework

The career-readiness platform. Learners build real projects in an in-browser IDE, and the agents author, test, and ship the curriculum they learn from.

on disk:khwajalabs-v2khwajalabskhwajalabs-v2-mobilekhwajalabs-lessonscontent-engine-kl

The Build IDE

An in-browser environment where learners ship portfolio projects through deep Socratic steps instead of copy-paste tutorials, with live assessments and an AI that reviews their work like a senior engineer would.

receipt: 4,500+ users, 45+ of them paying. 711 commits on the rebuild this year.

EdTechPaying customers

Curriculum written by agents, gated by humans

Give the system a project idea ('a fraud-detection platform') and it decomposes it into a beginner-to-capstone ladder, dedupes against the live catalog so nothing gets rebuilt, authors only the missing projects, and verifies them on a test environment before production.

receipt: Whole curriculum branches shipped this way; every project passes ingestion validation before a learner ever sees it.

Agent-authoredTest-gated

Demos and community on autopilot

A demo portal with product films recorded on an isolated local stack, plus Discord bots running weekly build challenges and project-of-the-week. Everything drafts first; a human approves before anything posts.

receipt: 3 films live; the community gets fed on schedule without me typing the posts.

DemosDraft-first bots

Ledger entry 03

The content factory

RemoteRee and everything downstream of it. The agents learned to write in my voice from my own corpus (journals, essays, a six-round live interview), and there's a spec with an impostor test to keep them honest.

on disk:content-engineyt-engineremotion-studiostory-enginecarousel-studiovideo-portal

The RemoteRee operator

A funnel-aware agent that runs a 3-reel-a-day cadence across seven content pillars, senses what performed, and plans the next day's slate. For a single reel, a chain of five agents runs: researcher, fact-checker (every claim needs an openable link or it gets cut), scriptwriter, SEO captioner, comedy pass.

receipt: 100K+ followers on the receiving end; 37 videos in the current engine.

100K+ followersFact-checked

Voice DNA

A portable spec of how I actually write, distilled from a close reading of my real corpus. It encodes the moves (one concrete image, concede-then-pivot, snark at the system never the reader) and, more importantly, a detector for the tells of an AI faking me.

receipt: Every public word the agents write passes a ten-point voice check, including this page.

Voice specImpostor test

A voice clone that narrates

My cloned voice narrates the long-form videos, with pacing math per voice: the clone reads at 156 words per minute, the stock Bangla narrator ambles at 83, and the subtitle and scene-length machinery compensates for each.

receipt: 1,119 voice clips generated and logged; full 8-to-12-minute documentaries narrated without a recording session.

Voice cloneLong-form

The $0 image department

Video frames render on the gaming laptop's GPU instead of a paid API: a fast 6-billion-parameter model for atmosphere shots, and a 20-billion-parameter model (squeezed onto a 14GB-RAM machine) reserved for any frame that needs legible on-screen text.

receipt: ~31-61 seconds per frame, $0 each, measured while the same GPU ran a TTS data job at 97% load.

Local GPU20B on a laptop

Story films, cast and scored

Writer-mode collaborations: I give a premise, the agent proposes scenes, we revise together, and it casts each character with a different cloned voice with per-line tone directions, then renders the film in Remotion.

receipt: 10-to-15-minute narrated films from a one-line premise.

FictionMulti-voice

Ledger entry 04

Hardware detours

Where the agents leave the browser. CAD, firmware, and a helium blimp, because apparently the ledger needed a line item that floats.

on disk:floating-swarmrobotics-forgebusiness-card-3dstl-parametricqr-stand

The floating swarm

A helium blimp swarm on an NVIDIA Jetson: firmware, a control dashboard, onboard vision, and the CAD, all drafted agent-first, then physics-tested against remeasured real-world numbers.

receipt: Airframe flying; control gains in retune after the physics said the sim was optimistic.

RoboticsJetson

Robotics forge

A bill-of-materials-to-build-to-verify-to-simulate pipeline, so hardware projects get the same gated CI discipline as software: nothing advances until the sim gate passes.

receipt: Builds blocked by a failing simulation, exactly as designed.

Sim gatePipeline

The parametric print shop

OpenSCAD everything: two-color business cards, QR stands, a bike tray, a conference badge with an e-stop pendant. Designed in code, printed on a Bambu P1S.

receipt: Real objects on real desks, including a badge worn at a real conference.

OpenSCAD3D printing

Ledger entry 05

Side quests

Small apps that each solved one real problem and then politely stayed out of the way.

on disk:runway-jobsmini-crmsongsmithprint-portalmiletracktx

Runway jobs

An agent that searches job boards and applies autonomously, including fishing the email verification codes out of Gmail mid-application when a form demands one.

receipt: Applications filed while I slept.

Automation

The mini CRM and six outreach streams

A lightweight CRM with a throttled email pipeline feeding six separate outreach streams, all governed by one standing rule: agents draft, a human presses send. Morning-ops passes sweep replies and bounces daily.

receipt: Six streams running for months; zero has ever sent without review.

CRMDraft-first

The event app and the shelf

A live event check-in app (still deployed) that parsed registration PDFs and handled the door, plus a songwriting tool, a 3D-print ordering portal, a mileage tracker, a budget copilot.

receipt: Shipped for a real event; the shelf tools still get used.

Live appTools

Ledger entry 06

The hijinx itself: 130 one-command procedures

The part people don't see. A 'skill' is a procedure an agent can run end to end from one command. Written once, they turn a solo founder into a staffed org. A sample of what single commands actually do here:

on disk:.claude/skills (130)shared/outreach-trackerspixel-portfolio

Tests that film their own demo

One command tests a running web app end to end with a browser, narrating and recording every scene as it goes, then stitches a captioned demo film plus an HTML report of what passed and what broke. The demo physically cannot drift from what the product does, because it is the test run.

receipt: All 13 published demo films were made this way.

/e2e-demo-report

A deploy pipeline for a whole fleet

One command builds, ships, and health-checks any app in the fleet to AWS; sibling commands spin services down to zero overnight, quote the monthly bill from the real billing API, and post deploy notes to Slack.

receipt: Every app in the portfolio deploys the same way; the AWS bill gets quoted, never guessed.

/deploy/costs

A job posting, decoded into a build

Paste a job posting and one command produces the lean actual requirements, a buildable portfolio project matched to that company, hooks for a video about it, and a link to the closest real project on the learning platform. It only links catalog entries that the live production API confirms exist.

receipt: Feeds the content engine and the KhwajaLabs catalog from the same paste.

/job-decode

A prospect demo, end to end

One command takes a company name, researches them, clones an interactive product demo customized to their facilities and catalog with working logins and a voice agent, then films and publishes a personalized narrated walkthrough.

receipt: From LinkedIn profile to a private, personalized live demo without a designer or an editor.

/will-call-prospect-demo

Memory, so mistakes cost once

238 files of institutional knowledge: which phone numbers must never be released, which auth route is untouchable, which published model silently fails to load its own weights, why agent prompts must stay under 2,000 characters. New sessions start with the scars pre-loaded.

receipt: The 11,000-character prompt that made an agent go mute on live calls is in there. It won't happen twice.

238 memories

The grid that runs it

Two laptops and a cloud on a private mesh: the Mac holds the truth, the gaming laptop does free GPU labor, AWS faces the public. Plus 53 parallel git worktrees with handoff files, a context archive so no work thread is ever lost, and a spend ledger with a line for every generated asset.

receipt: An unattended 99-call campaign day ran from the second laptop while the first kept building.

/victus-syncWorktree DB

Colophon

This page was researched, written, charted, and deployed by the same agents it describes. They backed up the old site first. We raised them right.

The stack behind this page: Next.js, Tailwind, and Framer Motion, deployed static to S3 behind CloudFront by a shell script the agents also wrote. The ledger closes here. It reopens tomorrow morning, on schedule.