Case study

Domain-Tuned Model for Children's Learning

A small, self-hosted language model tuned for age-appropriate tutoring — cheaper to run than a frontier API, and reviewable by educators before it reaches a child.

Client type
EdTech provider serving primary school classrooms (sample scenario)
Industry
SaaS
Duration
6 months

Sample contentSample engagement. Illustrative scenario used while the site is in build — not a named client project.

saas2026

Domain-Tuned Model for Children's Learning

  • Python
  • PyTorch
  • LLM integrations
  • AI agents

General-purpose models answer children the way they answer adults, and sending classroom conversations to a third-party API was unacceptable to the schools buying the product. Reading level, tone and safety all had to be controllable.

A compact open-weight model adapted with parameter-efficient fine-tuning on curated, age-appropriate material, with educators reviewing outputs in the loop and the whole thing deployed inside the client's own infrastructure.

Data residency objections from school buyers were removed, and per-conversation inference cost dropped far enough that the feature could be included in the standard plan.

  1. 01Curriculum-aligned dataset assembled with teachers, plus synthetic examples for gaps
  2. 02Parameter-efficient adaptation (LoRA) rather than full fine-tuning, to keep cost and footprint low
  3. 03Educator review loop scoring tone, reading level and factual accuracy
  4. 04Safety evaluations run as a gate on every model revision
  5. 05Self-hosted deployment so conversation data never leaves the client's environment

Smaller was the requirement, not the compromise

The brief was never “match a frontier model”. It was “be right about a narrow domain, run inside our own account, and cost little enough to give to every class”. A small adapted model met all three; a large API met none.

Start a conversation

Similar problem on your side?Let's talk.

Tell us what you are trying to build, fix or decide. If we are not the right team for it, we will say so and point you somewhere better.

Typical first step: a 30-minute call, then a short written assessment.