Avoca Edu
AI chatbot MVP · Education proposal · 2025
Who Avoca Edu is
Avoca Edu is a Vietnamese edtech platform from Avoca AI, combining K–12 subjects with AI and Robotics training. Their mission — "Learn AI, excel in the future" — is anchored by the FAT1000 program, which aims to identify and train 1,000 young AI and Robotics talents.
In 2025, Avoca Edu needed to submit a formal đề án (official project proposal) to the Sở Giáo dục và Đào tạo thành phố Đà Nẵng — the Da Nang Department of Education and Training. The proposal needed a working AI MVP behind it, not slides. That's where THECODEORIGIN came in.
What we built
Two production-facing AI agents, live in Avoca's own apps — one for teachers, one for learners.
Bloom Taxonomy assessment agent
In production on cms.avocaapp.com
A teacher-facing agent that ingests a student's English essay, pulls the relevant rubric and vocabulary context from the knowledge base, and returns a structured assessment mapped to Bloom's six levels — Remember, Understand, Apply, Analyze, Evaluate, Create. Every verdict is explainable: the teacher can see which rubric criteria the agent applied and why, so the assessment stands up to review.
Adaptive English tutor chatbot
In production on play.avocaapp.com
A student-facing chatbot that adapts its questions, difficulty, and vocabulary to each learner in real time. Grounded in the same curriculum knowledge base, it keeps every session on-syllabus and age-appropriate — pushing the learner without overwhelming them.
How we built it
The hard parts of an education AI aren't the model calls — they're the work around them. Getting the right knowledge into the system, framing the right question for each interaction, keeping outputs consistent enough to be trusted, and leaving the whole thing in a shape another team can pick up. Here's where the work actually lived.
Curating the knowledge base
Building the Vietnamese English curriculum into a retrieval-ready knowledge base.
Engineering the prompts
Calibrated prompts and structured outputs — the same essay earns the same Bloom level every time.
Designing the agent flows in Dify
Visual, inspectable flows — retrieve, reason, and branch for the edge cases.
Matching the model to the job
Gemini Flash 2.5 — fast and affordable at classroom scale, without giving up rubric quality.
Curating the knowledge base
The agents are only as sharp as what they can retrieve. We assembled the knowledge base from the Vietnamese English curriculum — textbooks, vocabulary lists, grammar references, rubric material — sourced from scattered, inconsistent formats. Everything was collected, cleaned, de-duplicated, and re-arranged into semantically meaningful chunks, each tagged with grade level, topic, and rubric dimension. When the agent retrieves, it gets the right slice, not an adjacent one. The AI looks sharp downstream because the data was made easy for it to read upstream.
Engineering the prompts
Ask a general model to score an essay by Bloom's Taxonomy out of the box and you get hand-wavy, inconsistent output. We iterated prompts with calibrated few-shot examples, explicit rubric anchors, and structured output contracts — so the same essay gets the same level across runs, and every verdict ships with the criteria it applied. The tutor side got the same treatment, tuned instead for in-session adaptation: harder vocabulary when the learner is coasting, simpler re-explanations when they're stuck.
Designing the agent flows in Dify
Both agents were orchestrated as Dify flows — retrieve, reason, structure the output, branch for edge cases (empty input, off-topic replies, unclear rubric level). Dify was a deliberate pick: the flow is visual and inspectable, which matters when the ultimate owner is another engineering team, not us.
Matching the model to the job
Gemini Flash 2.5 was a deliberate trade-off. Pro-tier models win on pure reasoning, but Flash is meaningfully cheaper and faster — which matters when the users are schoolkids chatting in real time and teachers batch-assessing essays. With a solid knowledge base and disciplined prompts, Flash was more than strong enough for the rubric.
Outcome
Tested, accepted, handed over
The MVP was accepted as part of Avoca Edu's submission to the Da Nang Department of Education and Training. At handover, the in-house engineering team at Avoca AI picked up the Dify flows, prompts, and knowledge base and continued evolving the product from there — exactly as intended.
Every choice we made — from the visual flow tool down to the metadata schema on the knowledge base — was made so another team could own it without reverse-engineering us. Both agents remain in use today on cms.avocaapp.com and play.avocaapp.com.
Exploring an AI MVP for education — or anything else?