Reviewing is slow
A thoughtful review of one notebook takes 20–40 minutes. Multiply that by a cohort and quality quietly collapses.
Aguacatito reads student notebooks and projects, then writes warm, specific, learning-first feedback — the kind a great mentor would leave. Designed by a UX & pedagogy practitioner, not bolted onto a generic chatbot.
No accounts. No upsell. Just a tool I'm building and a conversation about your classroom.
A thoughtful review of one notebook takes 20–40 minutes. Multiply that by a cohort and quality quietly collapses.
Most AI graders output vague praise or a score. A student can't learn from a number — they learn from a pointed, named observation.
Tired reviewers get terse. Students read terseness as “I failed.” The right tone is a teaching decision, not a footnote.
Each stage has one job and a clear contract. That's what keeps the feedback consistent, specific, and safe to put in front of a student.
Read the notebook: cells, sections, what ran, what errored, project type. Old review comments are stripped so re-reviews stay clean.
Deterministic checks first: did it execute? any errors? empty template? exposed credentials? No model needed for facts.
The model evaluates the judgment calls: is the analysis sound? do conclusions match the data? is the model evaluated honestly?
A rules engine turns findings into a verdict with a confidence level. Borderline cases are flagged for a human — not forced.
Feedback is generated against fixed templates and tone rules: a success note, a gentle warning, a thing-to-fix. Always in the student's language.
Each comment is inserted right after the section it's about, so the student reads it in context — not in a wall of notes at the end.
I spend my weeks between teaching and AI engineering, and I keep finding they're the same craft: curating the right information, in the right structure, at the right time, so understanding becomes possible. Aguacatito is that belief turned into a tool.
Comments support the next step instead of dumping the answer. Temporary structure that comes off as the learner grows — Vygotsky & Bruner.
Feedback targets what the student can almost do alone — the productive edge — not what's trivially right or hopelessly far.
Perkins' idea: you understand something when you can do something with it. Reviews check for performance, not recall of buzzwords.
Understanding isn't binary. Feedback names the level reached — recall, apply, analyze, create — and points at the next one.
“The teacher's lesson plan is the system prompt. The textbook is the retrieval corpus. The exam is the eval. Good teaching and good context engineering are the same discipline.”
— from my own notes, the idea Aguacatito is built on
Aguacatito's reviews run on Qwen — Alibaba's open model family — which gives strong, structured, multilingual judgment at a fraction of frontier-model cost. The pipeline speaks a plain LLM interface, so the same review can run on a local model on your own hardware or on a hosted API. Your students' work doesn't have to leave a machine you control.
I'm building this in the open. These are real figures from real runs — not projections, not a polished demo. Expect rough edges, and expect them to get better.
Run against actual student submissions — Jupyter notebooks and BI dashboard projects — not toy examples. That's where the rough edges show up, and where I fix them.
A human still approves every verdict. Coverage is uneven across project types. The model sometimes needs a second pass. I'd rather tell you that now than oversell it.
I take on a small number of consulting engagements: adapting Aguacatito to your rubric, your projects, and your students' language — or helping your team design feedback that actually teaches. Let's talk about what your reviews look like today.
Email hesusgc@gmail.com →A real reply from a real person — usually within a couple of days.