Skip to content
home
AI operations for accounting · hotels · law

When your best person leaves, the knowledge stays.

I build your business a memory: guests, rooms, clients, matters, prices, rules, and the way you actually handle each one.

It is built so a new hire can work from day one off the same source your agents use, and every answer an agent gives traces back to the document it came from.

Book a call (opens in a new tab)See the case studies
  • US Army veteran
  • Built on my own 50-room hotel first
  • Every claim carries an evidence label
01 / Platform

One core, five things it can do

Hermes is the part that knows your business. The five capabilities are the ways it reaches the outside world, and every one of them reads from the same map.

Agent coreHermes
What the agents read from — records · rules · how-we-do-it · source trace
Agent core

Hermes holds the map of your business and the rules you work by. Each capability is a door onto it. The chat window, the retrieval agent and the nightly job all answer from the same records, and every answer keeps the source it came from.

Select a capability to read what it fixes. Voice is in development and is described in future tense throughout.

  • Voice agents. In development — pilot on my own property first. Problem: After hours, during a rush, or when the line is busy, a new customer calls, rings out, and calls the next business on the list. What it will do: Pick up the calls your team would have missed. Say in the first sentence that it is an AI assistant, take the caller’s name, number, and what they need, then get a person on the line or book the callback. It would never touch your main line during staffed hours. What you would get: Every after-hours caller reaching something instead of voicemail, with a written summary of the call in your inbox and your calendar or CRM already updated.
  • Knowledge retrieval. Problem: The answer exists. It is in a contract, a rate sheet, a policy PDF, last year’s file. Finding it means someone stops what they are doing to go looking, or the caller gets told "I’ll check and get back to you." What it does: Reads your own documents and gives answers out of them, with the source page shown next to the answer. If the answer is not in your files, it says so instead of guessing. What you get: Your staff and your customers get the right answer in seconds, from your material, and you can click through to see where it came from.
  • Browser automation. Problem: Someone on your team logs into a portal every day to copy numbers into a spreadsheet, check a status, or re-key the same booking into a second system. What it does: Does that clicking and typing for you on a schedule, in the systems you already use, including the ones with no export button and no integration. What you get: The repeat login-and-copy jobs come off your staff’s plate and land in your sheet or your system, finished, before the day starts.
  • API automation. Problem: The booking is in one system, the invoice is in another, the customer record is in a third, and someone types it three times, sometimes wrongly. What it does: Connects the systems that offer a proper connection, so one event updates all of them at once, with a record of what changed. What you get: One entry, everywhere it needs to be, and a log you can check when something looks off.
  • Chatbots. Problem: People message your website, your Instagram, and your KakaoTalk at 11pm, and nobody answers until 9am. What it does: Answers from the same knowledge every other agent reads from, so no two channels contradict each other, and passes anything real to a person. What you get: Same-minute replies on every channel, and a running list of what people are actually asking you.
Proof

What is already running

These run on my own 50-room hotel. Each card carries an evidence label, says plainly that the property is mine and not a client engagement, and names the artifact I can open on the call.

  • CS-01Evidence label: Live / visible

    The morning housekeeping count arrives before anyone asks for it.

    • Hotel
    • Scheduled agent · PMS API · KakaoTalk · Slack
    • Running on the property since 4 Sep 2026

    MY OWN 50-ROOM PROPERTY — FOUNDER-OWNED, NOT A CLIENT ENGAGEMENT

    Situation
    My front desk opened the property-management system by hand to see which rooms still needed cleaning and which ones were being cleaned. Same lookup, several times a morning, done by whoever happened to be standing there.
    What I built
    A scheduled agent that signs into the PMS, reads the housekeeping board in one authenticated call, and writes a two-line Korean status message. It sends that message to the staff chat room and reports every run back to me. It deliberately does not assign rooms or write anything into the PMS. It reads and reports, nothing else.
    What it does now
    It runs on its own schedule through the morning on my own hotel, and each send leaves a delivery receipt I can open, so a missed run is visible rather than silent. It has not yet accumulated a 30-day reliability record. I have measured no labour or time figure, so this card carries none.

    artifact: Scheduled-run history plus a delivery receipt for every send. Not a public page.

    Not publicly viewable — walkthrough on the call
  • CS-02Evidence label: Live / visible

    Guests get answered in four languages, out of the house rules.

    • Hotel
    • Next.js · Convex · retrieval over an approved knowledge base
    • Public route re-verified 4 Sep 2026

    MY OWN 50-ROOM PROPERTY — FOUNDER-OWNED, NOT A CLIENT ENGAGEMENT

    Situation
    Guests ask the same handful of questions in Korean, English, Japanese and Chinese. Parking, late check-in, luggage, often at hours when nobody is at the desk. Before this they got whatever the person on shift remembered.
    What I built
    A guest chat page for the hotel that answers only from an approved Korean knowledge base and shows which record the answer came from. When the answer is not in the base it raises a request for a human instead of guessing. Behind it sits knowledge-base tooling for the people who own the answers.
    What it does now
    The public guest chat page is deployed and reachable. I re-checked it on 4 September 2026 before writing this card. The newer knowledge-base and conversation-review tooling is built and tested locally and is not deployed yet. There is no current production answer-quality measurement, so no accuracy figure appears here.

    artifact: The public chat URL below. Re-checked 2026-09-04: it served the hotel chat page with its Korean question field.

    Open the live page (opens in a new tab)
  • CS-03Evidence label: Live / visible

    Stay confirmations come out of a form instead of a spreadsheet.

    • Hotel
    • Vite admin app · Convex · PNG / PDF export
    • Live since Jul 2026

    MY OWN 50-ROOM PROPERTY — FOUNDER-OWNED, NOT A CLIENT ENGAGEMENT

    Situation
    Guests and partners ask for three specific Korean confirmation documents. Staff rebuilt them from Excel templates every time, which is where the typos and the mismatched formats came from.
    What I built
    A small admin web app that renders all three documents from one form, matched to the existing templates, and exports each as PNG or PDF. Entries save and can be searched back by reservation details, with a local fallback so a database outage does not stop the desk.
    What it does now
    It has been in routine use by my own staff since 11 July 2026 at the address below, with no code or deployment follow-up outstanding. I have not counted the documents it has produced, so there is no usage number on this card.

    artifact: The live admin tool at the URL below, in routine use by my own staff.

    Open the live page (opens in a new tab)
  • The full file

    4 builds are published, each one with what it does today and the artifact behind it.

    No build carries the Client paid label yet. The evidence key on that page says so.

    See all case studies

02 / Ontology

Your business’s shared memory

Every answer shows you where it came from.

Illustrative model — not a client record
Entity / Root

A 50-room independent hotel

Fields
  • Rooms it sells
  • People who work the desk
  • The rules it runs on
  • The systems it books into
Used by
  • Front Desk
  • Follow-Up
  • Back Office
  • Concierge
  • QA
Traceable to source

property record traces to room and rate map

  • A 50-room independent hotel has Guests. Read or written by Front Desk, Follow-Up, Concierge. Traceable to booking record, guest profile.
  • A 50-room independent hotel has Rooms & rates. Read or written by Front Desk, Back Office. Traceable to rate policy, OTA push rule.
  • A 50-room independent hotel has Invoices & folios. Read or written by Back Office, QA. Traceable to folio record, line item.
  • A 50-room independent hotel has Policies. Read or written by Front Desk, Concierge, QA. Traceable to house-rules document, parking section.
  • A 50-room independent hotel has SOPs. Read or written by Follow-Up, Back Office, QA. Traceable to rate-push SOP, OTA section.
  • A 50-room independent hotel has Staff handoff. Read or written by Follow-Up, Concierge, QA. Traceable to shift handover note, open items list.

Every edge is a stored relation. Every answer cites the record it came from.

03 / Your agent team

You don’t get a chatbot. You get a small team.

Five roles, one shared brain. Composed per industry, not per tool.

Illustrative model — not a client record

Roles marked running run on my own 50-room hotel (see case studies); roles marked designed are designed, not yet running.

  • RunningAG-01

    Front Desk

    Answers the guest messages nobody is free to take, in the guest's language, and raises a person when the answer is not in your material. Call handling is in development and would pilot on my own property first.

    Trigger
    A guest writes in on the site or a messaging channel, day or night.
    Hands to human
    Any refund, any rate exception, a booking request, or a guest who asks for the owner.
    Touches
    • Guests
    • Rooms & rates
    • Policies
  • DesignedAG-02

    Follow-Up

    Would chase the unconfirmed booking, the pre-arrival details and the review request, so nobody has to remember.

    Trigger
    A booking sits unconfirmed, or a stay ends.
    Hands to human
    A guest who answers a reminder with a complaint.
    Touches
    • Guests
    • SOPs
    • Staff handoff
  • RunningAG-03

    Back Office

    Reads the operational board on a schedule and posts the status where the team already looks, and builds the routine confirmation documents from your own records. Read-only on the systems of record: it reports, it does not write back.

    Trigger
    A schedule through the morning, plus every confirmation document a guest or partner asks for.
    Hands to human
    Anything the records disagree about. It reports the conflict; it does not decide it.
    Touches
    • Rooms & rates
    • Invoices & folios
    • SOPs
  • RunningAG-04

    Concierge

    Answers house rules, check-in times and parking out of your own material, with the source attached.

    Trigger
    A guest asks something the property has already written down.
    Hands to human
    A question your material does not answer.
    Touches
    • Policies
    • Guests
    • Staff handoff
  • DesignedAG-05

    QA

    Would check the other four against policy and put anything that went sideways in front of you before a guest complains.

    Trigger
    Every finished conversation and every scheduled pass.
    Hands to human
    Anything outside the written rule, before a guest sees it.
    Touches
    • Policies
    • SOPs
    • Invoices & folios
    • Staff handoff

    Checks the other four’s output against policy

Each one has a narrow job, a written boundary, and a rule for when it stops and gets a person.

04 / Use cases

Every job ends in a record you can check

The trigger, who works it, what it does, and what it leaves behind.

Illustrative model — not a client record

Steps marked running run on my own 50-room hotel (see case studies); steps marked designed are designed, not yet running.

  1. Step 1 · Trigger Running

    A guest writes in at 11:40pm

    The desk is empty. The message lands on the chat page, in the guest's own language, asking about parking and a late arrival.

  2. Step 2 · Agents Running

    Front Desk and Concierge work it

    Front Desk says in its first sentence that it is an AI assistant. Concierge answers both questions out of the approved house-rules material, and shows the record each answer came from.

  3. Step 3 · Action Running

    The booking itself goes to a person

    No room is held and nothing is written into the reservation system. The room request is raised as an open item for whoever opens the desk, with the whole thread attached.

  4. Step 4 · Trace record Running

    You read it in the morning

    The thread, the section of the house rules each answer came from, and the one item still waiting on a person.

    Example record — illustrative, not a client log
    source:
    house-rules document → parking section
    policy:
    answer only from approved material
    agent:
    Concierge, handed off by Front Desk
    human_review:
    required — the room request itself
05 / Continuity

When someone leaves, the map doesn’t leave with them.

Built so a new hire works from the same source your agents do, on day one.

Illustrative model — not a client record

Before: the knowledge is in people.

  • Shift 1
    • “knows which room types the channel oversells”
    • “remembers the late check-in workaround”
  • Shift 2
    • “knows the parking answer by heart”
    • “knows which nights the desk gets busy”
  • Shift 3
    • “knows the OTA rate quirk”
    • “remembers the Tuesday close”
    • “writes the handover from memory”

Shift 3 left. So did the process.

After: the knowledge is in the shared memory.

  • Shift 1
  • Shift 2
  • Shift 3
  • A 50-room independent hotel
  • Guests
  • Rooms & rates
  • Invoices & folios
  • Policies
  • SOPs
  • Staff handoff

Sourcerate-push SOP traces to OTA section

Shift 3 left. The process didn’t.

06 / Industries

Three verticals, written in their nouns

The map and the agent team are built out of your own nouns, so each vertical gets its own page. Only the hotel has published builds behind it today, and the other two cards say so.

  • HOT

    Hotels: See the hotel build

    A guest calls at 11pm, nobody is at the desk, and the same question gets a different answer depending on who is on shift.

    • Guest questions answered in four languages, out of your own house rules, with the record shown.
    • The morning housekeeping count sent to staff chat before anyone opens the PMS to look.
    • Confirmation documents built from one form instead of last year’s spreadsheet.
    Evidence label: Live / visible
  • ACC

    Accounting firms: See the accounting plan

    Half of busy season goes on asking the same clients for the same missing statement, while new work calls in behind it.

    • Would track what is outstanding per client, send the reminder, and log what arrives.
    • Would answer routine client questions from your engagement letters and prior-year files, with the page shown.
    • Would put a new-client call in front of a partner while the prospect is still deciding.
    No published build here yet — pattern proven on my own hotel
  • LAW

    Law firms: See the law firm plan

    Intake is the whole business, and the same matter details get re-keyed into three systems after it.

    • Would take the caller’s details and the matter type, then hand them to a person. It would not advise.
    • Would run a new name against your client and matter list before intake goes further.
    • Would find the clause in your own precedent bank and cite the document and page.
    No published build here yet — pattern proven on my own hotel

Not sure which one you are closest to? Compare all three side by side.

07 / How we work

Four steps, and you keep the map

Four steps. Each one ends with something written down that you keep, not a status update.

  1. 01

    Map the ontology

    One session on your records, your rules, and how you actually handle each case. I write it all down in one place, including what the system must never do.

    you get: the written map of your business

  2. 02

    Build one agent, prove it

    I pick the job that eats the most time and build a single agent for it. You watch it run on real work before anything else gets wired in.

    you get: one agent running, and its log

  3. 03

    Wire the team

    The other roles come in around the first one. Intake, follow-up, back office, research, and a check on what goes out, each one with the handoff to a person written into it.

    you get: the agent team, composed for your business

  4. 04

    Hand over with the trace log

    Every answer points back to the record it came from, and every action is logged. You keep the map, the logs, and the run-book, so your next hire reads the same source the agents do.

    you get: the map, the logs, and the run-book

What I promise

Every answer shows the record it came from

I build the map out of your own documents, run the system on my own business before it goes near yours, and hand over the trace log with it. If an agent cannot show you where an answer came from, it does not ship.

What I promise

  • The map is built from your own documents and records, not from a template.
  • Every answer an agent gives names the record it came from.
  • You get the trace log with the system, and it stays yours.

What I don't

  • A specific number of bookings, a revenue figure, or a result.
  • Anyone who guarantees you revenue on a system they haven't run in your business yet is guessing.

How you check

  • Open the record an answer cites and read it yourself.
  • Every claim on this site carries an evidence label.
  • Watch it run on work you pick before it touches anything live.
Changhee Lee · Founder
08 / Who builds this

I’m an operator who builds his own tools.

I’m an American and a US Army veteran. I run a 50-room hotel in Daegu, Korea, along with a coin laundry and a convenience store. Every system I have shipped runs on my own businesses before it runs on yours.

That is not a line for the website. It is how I find out what breaks, and it is why the parts that are not built yet say so on this page. If it isn’t a fit for your business, I’ll tell you on the first call.

  • US Army veteran
  • Runs a 50-room hotel
  • Builds on his own operation first
More about how I work
09 / Next step

Bring one workflow. We will map it on the call.

Tell me what your team repeats every week, and I will show you what the map for it looks like.

Book a call (opens in a new tab)Send a message instead
30 minutes · no deck · we look at your actual workflow