Can a free model on a 2017 CPU replace a paid API for reading receipts?
0.973F1 · $0
Qwen2.5-7B, 8.8 s per receipt. The 3B model stalled at 0.50, and I stopped tuning rather than fit the test.
ExperimentI'm Adit. I build the parts that keep a small business running: the booking page, the WhatsApp replies, the invoices. Every project below says plainly whether it's live, tested or still a prototype.
Demo of the tested JagaBukit booking flow. Not live with customers yet.
01 · What I build
For the owner who runs the business from a phone full of WhatsApp chats.
Price list, photos, the area you cover, and a button that opens WhatsApp with the booking already typed. Example: legacybarberhousecall.com.
Answers at 9pm, offers slots that are really open on your calendar, books the visit and reminds the customer the day before. It never quotes a price you didn't give it.
Invoices, follow-ups, receipts and a morning report, connected to the Gmail, Google Sheets or Notion you already use.
02 · Work
Client work, my own business, and builds waiting for their first customer. The badge on each one is where it actually stands.
Housecall barber covering villas, homes and hotels in Bali. One page; the booking goes out as a ready-written WhatsApp message.
ProofPaying client. Live since 23 Oct 2026.
My own body-alignment practice, villa and home visits only. English and Bahasa pages, two price tiers, WhatsApp booking with a tap counter.
ProofMy own business, not a client. Live since 29 Oct 2026.
Fill in the client and the session; it builds the PDF invoice and emails it from Gmail.
ProofRunning in my own practice. Started by hand, no AI step.
Free quiz that gives you an "AI founder archetype" and captures the lead.
ProofLive on Lovable. Lead capture tested 20 Aug 2026.
Books free villa health checks on the Bukit at any hour, offers only truly open slots, sends the reminder, never quotes a price.
ProofPasses 63 of 63 safety checks. Waiting on its domain and WhatsApp number.
Small Gmail, Notion and WhatsApp automations from a recipe library, rebuilt on Trigger.dev.
ProofDeployed and tested end to end. The WhatsApp ones still need their server.
An intake template for HVAC, law firms, med spas, real estate and four more: form in, Claude reads it, logged to a Sheet, WhatsApp reply, calendar slot.
ProofUnit-tested. Checks webhook signatures and drops duplicates. No client yet.
Paste your discovery-call notes, get a scoped and priced proposal.
ProofReal login, real proposal generated. Not launched.
A branded handbook plus a portal where a client can see their automations run.
ProofBuilt with Stripe billing. Not launched.
Reads a gym's Notion HQ at 7am, flags stuck tasks and writes the manager a plain-English brief.
ProofCode built. Membership numbers wait on the MindBody connection.
A portal for villa managers to see checks, reports and bookings in one place.
ProofDatabase live, portal unfinished.
Long video in, short clips out: transcribe, pick the moments, render.
ProofBeing built in phases.
03 · Lab
For the engineers reading this. The numbers are real, including the ones that went against my first guess.
Can a free model on a 2017 CPU replace a paid API for reading receipts?
0.973F1 · $0
Qwen2.5-7B, 8.8 s per receipt. The 3B model stalled at 0.50, and I stopped tuning rather than fit the test.
ExperimentDoes an AI need to see a chart, or is a caption enough?
0%→46.7%chart Qs
Captions alone answered none of the chart questions, even when they found the right chart more often.
ExperimentWhich search setup works best for legal research?
88%recall@1 · 6 ms
Plain keyword search (BM25) won. The reranker was about 475× slower.
ExperimentA research desk that refuses any claim it can't cite.
A critic agent rejects uncited answers. The eval runs in CI and fails the build below its threshold.
ExperimentSearch over my own skills and notes.
Retrieval and answer quality are scored separately, so a good answer can't hide a bad search.
ExperimentA voice assistant on my PC with Claude as the brain.
Talks back, opens apps, searches, remembers notes. Smoke-tested live on 18 Oct; 22 test files. A personal tool.
Runs locally04 · Sketchbook
Clickable mock-ups to test an idea before building it. There's no real AI behind them yet.
An AI class in Bahasa for total non-coders: build a site, an app, an automation, then an agent.
Edit a video by typing "remove the silences" or "add captions".
A pain-intake helper for therapists, filled in before the session.
A cooking app for beginners in Indonesian kitchens.
Bali sea and sky conditions on one map.
A proxy-shopping (jastip) app.
A one-page demand test for an AI receptionist for solo contractors.
Point your phone at a fish, get its name.
05 · How I work
Drafting is automatic. Sending, posting, paying and deploying always wait for a human to approve.
Before Claude edits code, Codex reviews the change and can block it. If the reviewer is down, edits stop.
The receptionist had to pass 63 of 63 checks first, including offering only real open slots and never quoting a price.
Every stack choice is logged with the reason and the tradeoff, so you can see why it was built that way.
06 · Contact
Send it on WhatsApp or book 15 minutes. If it's worth automating, I'll show you how. If it isn't, I'll say so.