Skip to content

Portfolio · Education & hospitality analytics

Two sectors, two questions,and no SQL in either answer.

Delivered for an IT services provider across several client engagements: a chatbot that turns a plain-English question into a query over a student information system, a forecasting model that tells a hotel operator what next month looks like, and — the part that outlasted both — the cleaning and feature-engineering standards the data science team adopted afterwards.

Two sectors, two questions, and neither of the people asking them wrote SQL.
01Sector
IT services provider — education and hospitality clients
02Lookup time
About 80% less for routine administrative queries
03Forecast accuracy
89%, validated out of time — five points over baseline
04Delivered
Chatbot, forecast model, executive dashboard, team standards
Act 01/ 033 figures

The question, in English

The SQL literacy barrier, removed rather than trained around.

Administrative staff at an education client needed answers out of a student information system, and had to ask somebody who could write a query. A T5 transformer generates the SQL instead. In front of it, an input classifier decides whether a question is one the system should answer at all, and topic modelling routes it to the right database — so a question that belongs nowhere is refused rather than guessed at. Routine lookup time fell about 80%.

The interface is a sentence. Everything behind it exists so that stays true.
Topic modelling picks the database before the query is written, rather than after it fails.
A question the system should not answer is refused — which is the harder half of asking in English.
  1. 01.01

    Generated, not templated

    A T5 transformer writes the query, so a question nobody anticipated still gets one.

  2. 01.02

    Relevance is checked first

    Out-of-scope questions are turned away rather than answered with a confident wrong number.

  3. 01.03

    About 80% less lookup time

    Measured on the routine administrative questions that used to become somebody else’s ticket.

Act 02/ 033 figures

The year ahead

Occupancy and demand, forecast and then actually used.

A hospitality client planned revenue and capacity on last year plus judgement. A Random Forest model over occupancy and booking demand replaced the judgement with a number: 89% accuracy validated on out-of-time data — not on a random split, which is the test a booking series will always flatter — and five percentage points better than the baseline it had to beat. It landed as an executive dashboard, and leadership took it into monthly revenue and capacity planning.

Validated out of time, because a booking series will flatter any random split you give it.
The model arrives as a dashboard, because a model nobody opens has not been delivered.
Adopted into the monthly planning cycle — the only proof a forecast ever really gets.
  1. 02.01

    Out-of-time validation

    Held-out future periods rather than a shuffled split, which is the only honest test of a forecast.

  2. 02.02

    Five points over baseline

    Stated against what the client was already doing, so the gain is the part that is actually new.

  3. 02.03

    It reached the meeting

    Revenue and capacity planning is where it is read, which is what made it worth building.

Act 03/ 032 figures

What the team kept

The models shipped; the standards stayed.

Across the same period the work spanned classification and NLP for several clients — topic modelling, named-entity recognition and object detection among them. The durable output was not any one of those models. It was a set of data-cleaning and feature-engineering standards, written down from what kept going wrong, and adopted across the data science team: the same preparation, the same assumptions checked in the same order, on every engagement after.

One preparation, applied the same way by everybody — which is what makes two results comparable.
Written down from what kept going wrong, rather than from what ought to be true.
  1. 03.01

    Several clients, one method

    Topic modelling, entity recognition and object detection, all prepared the same way.

  2. 03.02

    Cleaning is the model

    Most of the accuracy argument on these engagements was settled before a model was chosen.

  3. 03.03

    Adopted, not published

    The test of a standard is whether the next engagement uses it without being asked.

The last word

This shows how Famysys builds analytics for people who do not write queries: put the question in their own words, refuse the ones the data cannot answer, validate a forecast against time rather than against a shuffle, and leave the team a method rather than a model. If your reporting still runs through whoever knows the schema, that dependency is the thing to remove.

Built in

  • Natural language to SQL
  • T5 transformer
  • Relevance classification
  • Topic modelling
  • Demand forecasting
  • Random Forest
  • Executive BI
  • Feature engineering standards

All projects

Let's build what's next

Technology alone doesn't transform businesses.The right partnership does.

Modernizing systems, building a new product, or exploring AI-driven transformation — Famysys helps you move forward with confidence.

Schedule Executive SessionEmail Us Directly
  • SOC2 Type II Compliant
  • Strict Commercial NDA
  • Zero Lock-In Guarantee

Famysys

We Engineer Clarity

Scan to Connect

Instant digital business card & WhatsApp link