Charlie Tang Hoong

Metrics

What this assistant did, and what it cost. Client numbers can never be published; its own can.

100.0%
of costed projects were told not to build
1 of 1. An estimator that can only say yes is a contact form in disguise.

What happened to the questions

Last 30 days, straight from this Worker's own database.

  1. Asked4
  2. Answered by a model4
  3. Costed as a project1
  4. Advised against building1
Live from D1 No sampling 24 honesty rules llama-4-scout-17b-16e

Verdicts

Decided by rules in code, not by the model.

Do not build
  • Build00%
  • Fix process first00%
  • Do not build1100%

Answers per day

Zero days are real days with no questions.

Budget

Hard caps. At the limit it stops, never keeps spending.

Today4 / 120
  • Per IP30/hr
  • Per session8 turns
  • Free allowance10K/day≈69 answers

Honesty filter

Above 5% is a written reason to shut this down.

Answers blocked0.0% (0/4)
5%
  • Blocked before display0
  • Refused before the model0

Cost

Published prices × reported tokens — not an invoice.

  • Per answer$0.0016144 neurons
  • Last 30 days$0.0063
  • Worst case / month$2.39at 120/day
  • Tokens22.5K

Models

Workers AI only — no third-party key exists here.

  • llama-4-scout-17b-16e422.5K tok · 4.3s

Health

End to end, including tool calls and the filter.

  • p9510.3s
  • Sessions1
  • API degradations0said so instead of guessing
  • Stale knowledge base4used the snapshot

When this gets shut down

Written in advance, so retiring it is not an argument.

  • Fewer than 5 estimates in 60 days1 so far
  • Over $10 a month without real useclear
  • Filter blocking over 5%clear
  • Three months without a changeclear

The estimator still runs on a placeholder day rate of RM 1,800 — read its ranges as shape, not price. Every answer says so too.

How to read this page

Where the numbers come from. Every figure is a live query against this Worker's own D1 database, over the last 30 days. Nothing is sampled or rounded up.

Cost. Computed from Cloudflare's published Workers AI prices, read 2026-08-29, times the tokens each model reported — not from an invoice. Right order of magnitude, wrong number to quote as billed. Most of it is input: the whole site text goes into every prompt, which is the trade this design made instead of running a vector database.

Why a shutdown list exists. The failure mode for a demo like this is not that it breaks — it is that it quietly stops being true. The criteria were written before launch, so retiring it is a decision already made rather than one argued about later.

Why these numbers can be published at all. Two of the five cases on the main site carry results under NDA that can never be shown. This system is his own, so its numbers can be.