Meet NinjOrAItor: The AI Redefining How English Presentations are Assessed

For many university students, delivering a presentation in English is one of the most demanding parts of their academic journey. Whether explaining an engineering prototype, analysing a database system, or pitching a business idea, they are expected not only to demonstrate subject mastery but to communicate it clearly, confidently and persuasively.

Presentation skills, however, develop through practice and through timely, specific feedback. In traditional assessment models, feedback often arrives only after final submission, limiting opportunities for meaningful revision.

“Our students are mostly ESL or EAL learners, so they require more feedback and guidance to practise and achieve a better grade,” says Ms Christy CHAN, Senior Tutor at the Chan Feng Men-ling Chan Shuk-lin Language Centre (LC) at City University of Hong Kong (CityUHK) and Principal Co-Investigator of the NinjOrAItor project.

What students increasingly need, she explains, is an iterative system that enables them to practise, reflect and improve their presentations before the high-stakes final evaluation. The challenge is delivering that depth of feedback in a way that is efficient, scalable and sustainable within modern universities.

That challenge led Chan and her collaborators at The Hong Kong Polytechnic University to develop NinjOrAItor, a rubric-based hybrid intelligence platform that combines AI analytics with teacher and peer input.

Rather than replacing human judgment, the platform is designed to optimise it. “The program is more of an assistant for self-learning,” Chan emphasises.

Why Pure AI isn’t Enough
The team’s journey toward a hybrid model began with experimentation. Early versions relied more heavily on automated speech assessment, but that approach quickly revealed its limits.

“Purely AI-driven assessment tools only evaluate the surface level, for example pronunciation accuracy and fluency,” Chan says. “It won’t be enough. It’s not an ideal tool to use.”

Most AI systems assess measurable features: speech rate, number of vocal fillers, pronunciation accuracy. But effective communication requires nuance. “Sometimes you can use vocal fillers effectively for dramatic effect,” Chan explains. “But you can’t purely rely on the AI, because there are subtle differences.”

Bias is another concern. “Purely AI-powered speaking platforms will struggle with biases as well, the algorithm bias,” she says. Many speech recognition models are trained primarily on native-speaker data. “When we do the speech recognition, they might auto-correct your mistakes. This may mask the mistakes for you and fail to record the errors made, thus providing a less accurate assessment.”

AI also falls short on disciplinary context. “They lack the contextual knowledge to give a fair assessment on the students’ speech content,” Chan notes. For example, a database engineering presentation requires different rhetorical strategies than, say, a marketing pitch. Surface analytics alone cannot judge that distinction properly.

For the team, the conclusion to their experimentation was clear: AI could support English presentation assessment but it could not replace human expertise.

A Rubric-driven Hybrid Model
“We learn from our mistakes,” Chan says. “Hence we developed NinjOrAItor—a rubric-driven hybrid intelligence model integrating specialised ASR tools.”

The system works in layers. Students upload recorded presentations, which are transcribed and analysed by different AI agents. “We use Microsoft Azure for pronunciation, Deepgram for explicit filler word detection, and ChatGPT for commenting on the content,” she explains.

Crucially, educators upload their own rubrics. The AI generates draft feedback aligned to those criteria, but teachers remain in control.

“The teachers can upload their own rubrics,” says Chan. “The system automatically generates the AI comments, and teachers can edit those comments. Peers can also assess the same recording and give feedback.”

In this ecosystem, AI handles repetitive surface-level analysis at scale. Teachers focus on higher-order insights, conceptual clarity, disciplinary rigour and persuasive structure. Students engage in peer review and self-reflection.

“We’d like to believe the human–AI synergy is more like a teammate cross-checking each other’s work to achieve a faster and greater result,” says Chan.

Measurable Gains and a Shift in Mindset
Between February and June 2025, NinjOrAItor was implemented across seven courses in the Engineering and English disciplines at two universities. In total, 591 students made 1,186 submissions indicating iterative use.

In a study of 161 students who submitted both initial and final video assignments, the team measured a statistically significant overall improvement of 4.8 per cent, with particularly notable gains in pronunciation and vocal filler reduction.

Equally significant was the cultural shift in how students approached assessment. In another study involving around 60 students, participants reported feeling “more empowered to be in charge of their own learning”.

“They appreciate the fact that it’s lower stakes—they can do it at home before they go into a high-pressure situation,” Chan says.

Now, students can record, receive immediate AI feedback, reflect upon rubric criteria, and then incorporate teacher and peer input—all before submitting a final version.

As AI becomes more embedded in education, Chan emphasises safeguards. “First, data privacy: there’s no identifying information collected from the students,” she says. “Second is transparency. If we point out the error, we highlight it in the script and give a score explanation so they can choose to accept it or not.”

The project has already gained recognition and it’s been shortlisted for both the QS Reimagine Education Awards 2025 and the EDUtech Asia Awards 2025.

Yet for Chan, the broader significance lies in rethinking assessment culture itself. “It will transform the culture from more summative, high-stakes evaluation towards formative, iterative, learner-centred development,” she says. Feedback becomes continuous rather than episodic and students move from performing for grades to developing for growth.