Shikho Intelligence.

An always-on WhatsApp bot that answers leadership's data questions through guarded text-to-SQL. Numbers are computed in code - the model only narrates.

SEE HOW IT WAS DIRECTED →ALL SYSTEMS

INTERNAL SYSTEM · REAL INTERFACE, PROTECTED DATA · WALKTHROUGH AVAILABLE ON REQUEST

01 · THE PROBLEMWHAT WAS BROKEN

The company's numbers lived in dashboards nobody opened. Management lives in WhatsApp. A person spent 20 minutes every day screenshotting reports into the group.

02 · THE SYSTEMWHAT SHIPPED · WHO USES IT

It answers as a designed card, right in the group leadership already checks.

20 minutes of daily screenshotting, gone.

18managers served daily

20 minreclaimed per day

7+guardrail layers

SOURCE: His figures + decision log, July 2026

The morning scoreboard card - real design, synthetic numbers

The Control Center - target and schedule management, synthetic data

The end-of-day card with 6-channel breakdown - synthetic numbers

REAL UI · SYNTHETIC DATA · MORNING CARD · CONTROL CENTER · END-OF-DAY CARD

03 · UNDER THE HOODWHAT ACTUALLY HAPPENS

Pick a moment. Watch the calls the system makes that no one sees.

CODETHE MODELA CHECKTHE CLEVER BIT00/07

THE THINKING BEHIND EACH STEP

01It runs as its own always-on WhatsApp account, because the official API can't post to a group - a research finding, not an obvious build. Friday off, on the Dhaka week.

02Every post starts fresh and pulls the numbers live, so the card always shows today - a briefing built on a stale cached figure is worse than no briefing at all.

03Code owns every number - totals, pace, projection - because arithmetic has to be identical every run. A model asked to add or forecast can hand back a figure that's confidently wrong.

04By the time the model speaks, the numbers are already fixed - it only phrases them. It cannot invent a figure it never touched, so its worst mistake is an awkward sentence.

05Fonts and logo bake into the image so it draws pixel-identical anywhere - and it lands as a designed card, not a wall of text, because leadership reads the group in glances.

06It retries with backoff, per group, because a flaky WhatsApp minute must never silently drop a briefing - the failure mode to design against isn't a wrong post, it's a missing one.

07A milestone card is a nice-to-have; the briefing is load-bearing - so it's sandboxed to fail alone. A celebration bug never touches the numbers, and a watermark fires each rung exactly once.THE CLEVER BIT

EVERY STEP MINED FROM THE REAL BUILD · NOTHING INVENTED

04 · WHY NOT ONE PROMPTWHAT A ONE-SHOT WOULD HAVE GOT WRONG

A one-prompt WhatsApp bot demos well. This one has to be right in front of leadership every morning.

The platform itself was a research finding. The official WhatsApp API cannot post to a group. A dedicated always-on account, two linked devices kept apart, came from deep research.
Nothing here runs on a default. Thirty-five numbered decisions govern it. Which model writes the SQL was measured - the brand-name favourite silently broke date filters.
It is grounded in data no model has seen. A 786-line data dictionary - 194 tables, forty gotchas - feeds the bot.
It was battle-tested before leadership saw it. A 56-question bilingual harness caught six bugs - including a jailbreak trying to pull personal data.

05 · THE DIRECTIONWHERE THE HUMAN WAS

THE FORK · WHAT THE AI PITCHED VS WHAT I SENT

Mid-June. He asks for May's numbers; the bot answers confidently with 2023 data. He pastes the screenshot: "What happened here, how did it get 2023?" The model's instinct: patch the prompt.

THE MODEL'S POST-MORTEM, TWO DAYS LATER

Most of the time it's right. But I found cases where it confidently gives a wrong number and nothing catches it. [...] Every fix this session was "add another sentence to the prompt / another veto regex." That approach has a ceiling, and I hit it - I found 7 fresh failures in under an hour of the *current* code. The architecture, not the prompt, is the limiter.

WHAT I SENT

Based on the data dictionary file we have - would it make sense to go through that and maybe basically update the current system we have for this bot to better understand questions and find out how to run this?

The fix that shipped was structural, not another prompt patch. The date window, every number, the definitions and a second-model review all moved into code.

SESSION TRANSCRIPT · JUNE 2026 · BOTH SIDES QUOTED EXACTLY · [...] MARKS OMITTED SPEECH

A leadership bot cannot be "usually right". Everything dangerous moved out of the model's hands.

35 logged decisions56-question regression harnesshard PII wall226 real questions answeredHOW IT WAS RUN, NOT HOW IT WORKS

NEXT SYSTEM AI Content Studio →

▸