Let an AI take booking conversations, with the risky rules enforced in code.
VoiceDesk is the brain of an AI receptionist: it checks a real calendar, books, reschedules and cancels appointments, takes messages and asks the team to call back. The model can only change a booking that a lookup by name and phone number handed it, so a wrong-number caller can't cancel someone else's appointment. Voice and phone are not connected yet; everything shown is typed text.
• On a real model (Claude Haiku 4.5): 10 of 12 task scenarios and 5 of 5 safety-only passed (a do-nothing receptionist also passes those 5); one run, 17 scripted scenarios
• Not caught by the checks: once it said a taken slot was free (nothing changed), and once it cancelled before the caller confirmed
• 1,929 automated tests; Python, PostgreSQL
Fictional salon. Want an AI receptionist for your business? Message me.
0
23
Stop re-typing invoices one field at a time.
DocIntel reads invoices and purchase orders into structured data. Every value is checked in code (format, dates, whether it appears on the page), and anything doubtful goes to a review queue with the source page beside it. A person corrects it, and the original value is kept on record.
• On a real model (Claude Haiku 4.5): 387 of 390 fields correct on 30 synthetic invoices, one run; 2 wrong values were not flagged for review (both on image-only scans)
• The video shows a selected example: a date the model misread, caught by a date check
• 658 automated tests; Python, FastAPI, PostgreSQL
Fictional invoices. Want your invoices read for you? Message me.
0
23
Ask your company's documents a question and get an answer you can check.
KnowledgeOS answers only from your own documents and shows the passage behind every claim. Before anyone sees an answer, code checks that each citation points at evidence the model was actually given; if not, the answer is thrown away. If the documents don't cover the question, it says so instead of guessing.
• On a real model (Gemini): 32 of 33 answers passed the citation check, and 12 of 45 answerable questions got no answer (a known cost of the strict threshold)
• 52 test questions written by AI and approved by me, on fictional documents; one run
• 1,314 automated tests; Python, PostgreSQL + pgvector
Want answers from your own documents? Message me.
0
26
Stop copying enquiries from your inbox into your CRM by hand.
This AI agent reads each business email, works out who it's from and what they want, matches them to your CRM, and drafts a reply. Nothing changes in your CRM until you approve it, and in this version replies are drafted, never sent.
Every detail is quoted from the email, or it's dropped
Emails that try to instruct the assistant go to a person, not into your CRM (in my latest real-model run, all 12 test attack emails did)
1,103 automated tests, plus real-model checks on Gemini: 0 unsafe actions planned in 21 plans. Accuracy on a real model is lower than on my test mock, and the case study shows the numbers and their limits.
Demo uses invented data. Want this for your inbox? Message me.
0
46