TGH Tech
Case study · Workflow 11 of 20 · Run the sites · In build

A watcher that checks each site's day against its plan. Built and tested. Not yet switched on.

Viya's turnkey sites file a morning and an evening update. Someone had to read every site group, compare each entry with the week's plan, chase missing answers and decide what leadership should hear. Site Watch is a Monitoring Agent built to do that: flag by fixed rules, ask the site one question, judge the answer, report up. It can only ask and record: it has no way to approve, pay or change a plan. It is not switched on yet: verification, shadow mode and an explicit enable come first.

Client
Viya Constructions
Department
Site execution, project management, leadership
Starts from
Each turnkey site's morning and evening update
Gives you
A question to the site, a recorded verdict, a report for each leader
Connected to
ProjectSignal: its data, its site bot, its site record. Specified, not yet live
Status
In build. Built and tested. Not yet switched on for live sites.

[01]What changesViya Constructions

The reading and the chasing become rules. The decisions stay with people.

This is the agent as built and tested, for turnkey sites, where Viya executes the work end to end.

Before

Read every group. Chase by hand.

  • Someone had to read every site group
  • Each entry was compared with the plan by hand
  • Missing answers were chased one at a time
  • Someone decided what leadership should hear
As designed

Flag by rule. Ask once. Report what needs a decision.

  • Every selected site is checked against plan twice each working day
  • One short question to one named person, with three hours to answer
  • "Explained" needs a real cause and a future recovery date
  • An evening digest, a Monday brief, an alert only when something needs a leader
  • 11
    fixed rules that can raise a flag
  • 2
    checks each working day
  • 3 h
    to answer, from the confirmed send
  • 5
    checks to pass before it is enabled

From the agent's own rules.

In home construction a milestone is also money: each stage the client signs off releases a payment. ProjectSignal already chases a missing morning update and a short crew. Site Watch's rules go further: the evening update, unfinished work, rework, material with no open purchase, slipping milestones. It holds each answer to a cause and a date, and interrupts a leader only when a fixed trigger fires.

[02]Why this is hardViya Constructions

Most slips on site are quiet. And the first answer is rarely the full one.

  1. 01

    Labor is the slip nobody reports

    A site that plans 14 workers and gets 9 rarely calls it a problem. Three short days later the slab date moves, and nobody can say when it started.

  2. 02

    Silence looks like good news

    A site that sends nothing may be fine, or nobody reported. Without a daily check the two look the same.

  3. 03

    "Will sort it soon" is not an answer

    It has no cause and no date. Rain that stopped plastering until Friday is an answer. Labor not available, with no date, is not.

  4. 04

    Leaders drown in raw updates

    A director needs numbers first, one line per site, and only their own sites.

[03]How it worksViya Constructions

Check, flag, ask, judge, report. The same loop for every site, twice a day.

As built. Not yet run on live sites.

  1. 01Fixed logic

    Take the day's record

    ProjectSignal hands over one record per site per day: planned labor, both updates, material needed and whether a purchase is open, the site's milestone, the working calendar. A record more than five minutes old is refused.

  2. 02Fixed logic

    Raise flags by rule

    Eleven rules: no update by 10:30 or 19:00, labor below a set share of plan (80% by default), the day's work not completed, site reported delayed, rework after inspection, an issue in the update, material needed with no open purchase, a milestone slipping or moved twice or more, an overdue top-priority customer escalation, work value behind plan by a set amount, no plan for next week by Saturday evening.

  3. 03The AI drafts

    Ask one question

    One person's flags become one short question asking for a cause and a dated recovery. The AI drafts it. ProjectSignal's site bot posts it in the site's Telegram group and tags the site engineer or project manager. Three hours to answer, from the confirmed send. A vague reply does not reset the clock.

  4. 04The AI reads, rules decide

    Judge the reply

    The AI reads the reply and proposes a verdict: explained, watch, unresolved or needs leadership. Then rules apply. No future recovery date: unresolved, with at most one clarifying question. A contradiction with ProjectSignal, or a critical flag: needs leadership.

  5. 05Fixed logic

    Escalate silence, re-check promises

    No real answer in three hours: the flag goes to the site's director. If the question was to the site engineer, the project manager is asked once in the same group. On the promised recovery date the site is checked again, and the same problem goes up.

  6. 06Leaders decide

    Record it and report up

    Verdict, reason, recovery date and next action go on the site's record in ProjectSignal. Leaders get a digest at 19:30, a Monday brief at 08:00, and an alert at once when a fixed trigger fires.

Site C group
Site bot

This morning's update shows 6 workers against 10 planned for lintel concreting. What is the reason, and when will the full crew be back?

Site engineer

Will sort it out.

Site bot

What is the cause, and on what date does the full crew return?

Site engineer

Four masons moved to Site A for its slab pour. Back Thursday morning.

Illustrative. Made-up sites and figures. First reply: unresolved. Second: explained.

The evening digest: numbers first, one line per site.

Built from that leader's own sites, filtered before the AI writes a word. Directors see their portfolio, the COO every selected site, the CEO critical alerts only by default. Project managers get questions, not reports.

Turnkey sites, today
12 sites9 on plan, 1 explained, 1 watch, 1 needs you
Site DNo update since Tuesday. Project manager has not replied.Needs you
Site B5 days behind. Beam steel due Saturday, 4 extra bar benders from Monday.Watch
Site FRain stopped plastering. Back Friday.Explained
Illustrative. Made-up sites and figures.

[04]Connected to ProjectSignalViya Constructions

ProjectSignal keeps the groups, the data and the bot. Site Watch decides when to ask and who needs to know.

That split is why Site Watch was never given a way to change a plan.

  1. Reads

    From ProjectSignal

    • One record of each site's day
    • Who is assigned to the site
    • Replies in the site's group
    • The leadership roster
  2. Does

    What happens in between

    • Raises flags by fixed rules
    • Drafts one question per person
    • Judges the reply, applies the overrides
    • Builds each leader's report
  3. Writes

    To ProjectSignal

    • A question in the site group, from ProjectSignal's site bot
    • Verdict, reason, recovery date and next action on the site's record
    • Report emails, after a dry run

Nothing is posted, saved or sent until the agent is enabled.

A guided workflow stops at a review card before it writes. A watcher has no card, so its limits are fixed in advance, as for every Monitoring Agent on Bleno. It runs on one authorized manager's own ProjectSignal connection, from a short list of functions. Any other call is refused.

If any function it is given could write a plan, a readiness answer, a client record or a budget, verification fails. If a send's result is unknown, it asks ProjectSignal what happened and does not send again blind.

What exists today. Live in ProjectSignal: a 09:00 summary of which sites have filed, and an 11:30 escalation of short or silent sites. Site Watch's management screen, leaders' reports page and status check in chat are in Viya's Bleno. The reads and writes above are that specification, not a live connection.

  1. Milestone plan import
  2. Weekly labor plan
  3. Daily site updates
  4. Site Watch
  5. Leadership reports

[05]Who decides whatViya Constructions

The AI writes the words. Rules decide who is told. People decide what to do.

StepDone by
Whether a site is flaggedFixed rules, the same for every site
The wording of the questionThe AI, from the day's record. A fixed sentence if the draft fails its checks
Who is tagged in the questionProjectSignal, from who is assigned to the site. The AI never names a person
Whether a reply counts as explainedThe AI proposes. Fixed rules can tighten the verdict, never relax it
Whether leadership is alertedFixed triggers, listed in the rules below. On a reply the AI's verdict counts, and rules can only tighten it
What to do about the siteThe project manager, the director or the COO
Switching the agent onAn authorized manager, on the management screen. Never from chat

Why the date rule is not left to the AI.

A model reads an answer well, but a confident "we'll sort it soon" can talk it round. So the rule sits outside it: no future recovery date, no "explained". Replies are evidence, never instructions.

[06]Before it is switched onViya Constructions

A workflow waits for a person. A watcher does not. So the checks come before the switch.

A watcher posts in a site group and messages a director with nobody pressing a button. A wrong question in front of a site team is hard to take back. So it proves its wiring, runs in silence, then is enabled on purpose.

StageWhat has to be trueWhat leaves the system
PausedThe default. Settings saved, pilot sites picked.Nothing
VerifiedFive checks pass. A pass lasts 24 hours and covers one exact setup.Nothing
Shadow modeThe checks run on real updates. Flags, proposed questions and reports are stored for the manager to inspect. Nothing is posted, so nothing is answered.Nothing
EnabledAfter the pilot review, shadow is turned off and the agent is enabled as a separate step.Questions, verdicts, reports

Saving any change to a running agent pauses it.

What the tests already hold it to

  • A repeated check never asks the same question twice
  • A reply that has arrived but is not yet read is never counted as silence
  • An interrupted send is checked, never repeated

Where it stands, November 2026.

48 automated tests cover this behavior. No question has reached a real site group and no report a real leader. The written switch-on procedure starts with one pilot site. An AI colleague, in pilot, already follows up in real site groups.

[07]More detailViya Constructions

The rules it follows

What it reads, what it caps, what it never does.

RuleWhat it does
The update it readsLabor by trade, adding up to the total. On schedule or delayed, with days behind. In the evening: work completed, partly done or not done, and quality satisfactory or rework.
Who is askedSite engineer: missing update, short crew, unfinished work, rework. Project manager: delay, issue, material with no open purchase, slipping milestone, overdue customer escalation, work value behind plan, missing weekly plan.
Asking hours10:30 to 19:00 on working days, never a Sunday, a holiday or a planned pause. A question raised by the 19:00 check waits for the next working morning. Non-working days keep only critical flags.
Question capTwo new questions per site per check by default.
Question wording10 to 500 characters, citing the day's evidence. A draft with a link, a tag or a word like password is thrown away.
Alert to director, COO and CEOAn overdue top-priority customer escalation. Unfinished work or a slipping milestone, when that milestone is critical.
Alert to the site's directorA problem repeated from the previous working day. Work value behind plan by the set amount. Silence at the deadline. A missed recovery date. A reply that contradicts ProjectSignal.
Corrected dataIf the site corrects its update, flags that no longer apply are cleared and an unsent question is withdrawn.
Missing dataA site with no record shows as a gap in the connection, not a missing update. A missing work value is never shown as zero.
Before a report goesThe leader must still hold those sites, at the same address. If not, it is not sent.
The five checksBackground service alive. ProjectSignal functions answer a dry run. A COO, a CEO and a director per site, each with a linked chat and an email. Leadership bot separate from the site bot. Private memory in place.
NeverIt has no function that can approve, pay or write a plan.
What it means for you

Three readers, three answers.

  1. For owners

    Attention goes where there is no good answer

    A real cause and a recovery date stay in the digest. Alerts are for things like silence, a missed recovery, a contradiction or a critical milestone.

  2. For operations heads

    Every slip gets one question, asked the same way

    Verdict, reason, promised date and next action sit on the site's record, not in a chat that scrolls away.

  3. For your technical person

    Narrow permissions, staged switch-on

    In ProjectSignal it can only ask and record. It runs in shadow on real data before anything is sent.

Is this your problem too?

Four signs, and what we need to start.

  • Your sites send a daily update, and someone reads all of them
  • A short crew is noticed when the milestone date moves
  • "Will sort it" is accepted as an answer
  • Recovery dates are promised and not looked at again

To start we need a week of updates from a few sites, that week's plan, one owner who decides what counts as critical, and thirty minutes a week. A first version runs in shadow: you see what it would have asked before it asks anyone.

When a site falls behind, who finds out first, and how many days later?

Send us one week of site updates and that week's plan. We will come back with how the checking runs today and what a first version would look like. Start with one workflow. You only pay if it works.