Welcome to Braedyn Thompson's portfolio!
See the roles I'm targeting →

Now playing
Berkeley Lab SeismicSoCal BearLM
INTERNSHIP · BACKEND · AI

Kigumi Group

Backend Software Engineer Intern in Hong Kong on KiguLab, a digital wellbeing and AI ethics curriculum for students. I hardened the AI assistant pipeline behind it and pushed test coverage to 98%.

Jun 2025 – Sep 2025Hong KongBackend · Testing · LLM pipelines
Visit KiguLabThe Kigumi Group
backend code coverage90%before98%after

Overview01

98%Code coverageup from 90%
-80%Untested code10% → 2% uncovered
2Failure classes fixedconnections · query formulation
2Model familiesGPT + image generation

An AI assistant for kids' education has to work every time. A student who gets an error or a garbled answer usually doesn't try again.

The Kigumi Group builds KiguLab, a digital wellbeing and AI-ethics curriculum for ages 8–21. It covers digital wellbeing, critical thinking, ethical AI companionship, cyber resilience and privacy, through short "edutainment" lessons in English, Thai and Chinese, with AI tutors that teach alongside human coaches. I worked on the backend behind the AI assistant.

  • 01

    Increased code coverage from 90% to 98% and resolved recurring AI assistant pipeline failures caused by unstable backend connections and malformed query formulation to the image-generation and GPT models. I did this by writing unit and integration tests and diagnosing issues through targeted debugging, which improved reliability and response consistency.

KiguLab landing page
ScreenshotThe KiguLab landing page: digital wellbeing and AI-ethics lessons for ages 8–21.

Fixing the AI assistant pipeline02

The assistant takes a learner's message, builds a query, and sends it to either a GPT model for text or an image-generation model for pictures. Failures kept coming back, and they came from two different places:

Two failure classes that looked alike from outside, with two different fixes.

AI assistant request path (simplified illustration)LEARNERasks the assistantBACKENDAI assistant serviceQUERY BUILDprompt formulationGPT MODELtext answersIMAGE MODELimage generation121 Unstable backend connections dropped or stalled model calls2 Malformed query formulation sent bad requests to the GPT and image models
IllustrationSimplified illustration of the request path, not the production architecture. The red markers are the two failure classes I tracked down.

1 · Unstable backend connections

Some model calls dropped or stalled because the connection between the backend and the model services wasn't reliable. To the learner this looked random: the same question would work once and fail the next time.

2 · Malformed query formulation

Other failures came from the request itself. Queries to the GPT and image models were sometimes put together wrongly, so the model got a bad request and either errored or answered inconsistently.

Why did the assistant only fail some of the time?

Two separate causes: unstable backend connections, and malformed queries sent to the GPT and image models.

Same symptom, different bugs. Sneaky.

How I fixed them

  • Targeted debugging to separate the two failure classes, since they looked alike from the outside but needed different fixes.
  • Integration tests that exercise the real request path to the models, so connection and request-shape problems show up in testing instead of in front of a student.
  • Unit tests on how queries are formed, so a malformed request is caught by a test before it ships.

The result was fewer pipeline failures and more consistent responses from the assistant.

Kigu, the KiguLab AI Buddy, in a lesson
ScreenshotKigu, KiguLab's AI Buddy, inside a lesson: it opens the course, asks a reflection question and takes the student's answer. This is the assistant whose backend pipeline I made reliable.

This is the assistant whose pipeline I fixed.

From 90% to 98% coverage03

Coverage was already 90%, and the last 10% is usually the hardest: error paths, edge cases, and code that talks to external services. That's also where the pipeline bugs were.

Backend test coverage0%25%50%75%100%BeforeAfter my test suiteCode coverageCode coverage — Before: 90%90%Code coverage — After my test suite: 98%98%Uncovered codeUncovered code — Before: 10%10%Uncovered code — After my test suite: 2%2%Uncovered code shrank from 10% to 2%: four fifths of the untested surface closed.
FigureCoverage went from 90% to 98%. Put the other way, the untested part of the backend shrank by 80%.

Coverage from 90% to 98%: the hardest 8%.

  • Unit tests for backend logic in isolation, including how queries to the models are built.
  • Integration tests for the joins between components, where unstable connections and bad requests actually fail.
  • The failures I diagnosed were pinned down by tests, which guards against regressions.