workalike

workalike is the agent that learns the rules nobody wrote down. See it in your browser, about a minute, no install: https://flyingrang2.github.io/workalike/ Check the numbers, no GPU: https://flyingrang2.github.io/workalike/evidence.html The page plays one customer request handled three ways on the same model and GPU, from the recorded runs, sped up: 32 seconds and 580 tokens with the LoRA, 329 and 6,341 out of the box, 425 and 8,130 reading the rules. Built solo during the hackathon: the first commit is from October 27, 2026, and there is no earlier codebase. Runs on Meta's open-weight models: Muse Glimmer, self-hosted with vLLM and trained with QLoRA on one rented NVIDIA L40S 48 GB, with a hosted version on Muse Spark Open through Baseten (`meta-models/muse-spark-open-1026`). It is agentic: it works a ticket end to end with the desk's tools and stops at a person's approval. I run a one-person startup, and the work is different every time. When I hand a job to someone else, it rarely goes the way I want, so I end up doing it myself. I wanted someone who works like me when I'm not there, and that is why I built workalike. Then I timed myself against it, by hand, on one ticket from the test: 3 minutes 23 seconds, and my reply had no quote. The LoRA took 31 seconds and got it right, and across all 48 test runs it averaged 29. My draft also slipped past the desk's checks, so we fixed the checks. When the person who knows a job goes on vacation, the tools stay and the judgment leaves. A manual is slow to write and slow to read, and the thing a customer needs often was never written down. The first person this helps is the colleague covering the desk: they get a draft done the absent person's way, and they approve it. We built this on a sample quote desk for a small print shop. A person works as usual and a switch decides what is recorded. Muse Glimmer reads the records, writes the rules it finds, and asks only what the records could not settle. The model practises on new tickets from those rules, and the clean, correct work trains a per-person LoRA with the rule book taken out. With that adapter, Muse Glimmer handles new tickets with none of the rule book in its prompt. The rule book stays the source of truth: change a rule and the adapter rests until it has learned the change. Every reply stops as a draft, and nothing goes out until a person approves. The desk's own checks read each draft against the desk's numbers; first drafts that told a customer the shop's floor price were held back. For our test, 60 records came from a scripted worker, one person answered 29 open points, and our practice script connected practice and training. The product's own learning round is a separate method. One person, the founder, spent about 10 to 15 minutes, by their own estimate, answering 29 questions. After about two hours and about $8 of GPU time (about $10 with one step rerun), Muse Glimmer was working by those rules with none in its prompt. We measured it on 24 unseen tickets, with answers written and hashed before any run, twice each at temperature 0 through the product's own agent loop. It got 41 of 48 right: 37 out of the box, 45 reading the written rules on every ticket. Dangerous drafts were 7, 0 and 1. In a separate run, one arm alone on the same GPU, it worked about five times the tickets an hour (485 against 100 reading the rules, or 425 against 96 counting only the right ones). In the main test, p95 came down from 456 seconds to 68. On three tickets with no ship date, twice each, the model out of the box quoted a date nobody gave in 6 of 6 runs, and with the LoRA it asked in 5 of 6. When we ran the product's own learning round on a real GPU, the new adapter failed its check and the product kept it off. Every run can be read on the evidence page, one command recomputes the numbers, and the limits are in the README. Next is a real shop's desk, then one computer in the office and one workalike for every person, with the data staying in the building. We have served nine adapters from one server on one GPU, as a serving test. The code and the sample adapter stay free. We plan to charge small shops for what they can't do themselves: turning one job into checked rules and a trained adapter, on their own machine or on one we host, and keeping it current as the rules change. Our first test is a fixed-price pilot with one real shop.