Skip to main content

Personal AI Research Mentor

Human Development Technology software

Intelligence System helping you grow as a researcher

You want to learn to do real research. Like any skill, it can be learned from a mentor. AI Research Mentor provides that guidance: it asks the right questions, challenges your thinking, and helps you design rigorous experiments. It adapts to where you are, so you have mentor-level expertise whenever you need it.

This is the first source-available research guidance system designed by scientists specifically for student researchers.

The vision behind Personal AI Research Mentor is set out in the paper below, which presents the system as the first implementation of Human Development Technology: educational software where the teaching methodology itself is open to inspection and anyone can reproduce the conditions to study the system independently.

Papers

See It in Action

The Question Workshop, start to finish: one word becomes a research question you can actually work on.

What the Software Does

The mentor does its reading, searching and checking in the background, to work out where you are and what to ask you next. The research itself stays yours to do.

A mentor you can talk to

An open conversation about whatever you are working on right now. It remembers your project between sessions, so you pick up where you left off instead of explaining yourself again.

Under the hood
A 21-node agent, built on LangGraph, decides what to ask next. Memory is scoped to one project, so two projects never bleed into each other, and everything in and out passes a safety review.

Pick the scientist whose method you want

Your choice shifts the mentor's priorities, its approach, and the tone of the conversation. Work in the style of Newton, Galileo, Leonardo da Vinci, Mendeleev, Pasteur, Darwin, Ada Lovelace, Sofia Kovalevskaya, Ibn al-Haytham, Aryabhata or Zhang Heng.

Under the hood
Thirty-three in all: twenty-eight historical, across the European, Islamic, Indian and Chinese traditions, and five contemporary approaches. The steps it suggests and their order are inferred from how that scientist actually investigated; their way of explaining, and the kind of example and analogy they reached for, carries into how the mentor writes. Not their period language, though, and it never announces whose method it is using. The teaching approach underneath is modern mentoring pedagogy, so the historical method rides on top of what is known to work with students. Newton is the default.

The Question Workshop

Nine different routes from a vague interest to a research question you can actually work on. Most return a ranked set of candidates to choose between; the hypothesis route instead walks you through designing one, stage by stage.

Under the hood
The nine are phenomenon, claims, gaps, theory, modeling, forge, questioned, retractions and hypothesis. Each is a different starting point rather than nine passes of the same one. The hypothesis route runs as a staged agent: variables, literature, hypothesis formation, dataset search, confounds, operationalisation, analysis plan and resources.

Retracted work as a source of questions

Papers get withdrawn, and the gaps they leave reopen. One workshop route works from the Retraction Watch record: what failed, what it did to the field, and which questions are live again because of it.

Under the hood
Retraction Watch publishes a daily record of retractions, expressions of concern and corrections, each with a curated reason. It is downloaded, embedded and searched on your own machine, refreshed weekly, and needs no API key. The pipeline runs five stages and returns three to five ranked questions with feasibility scores.

Grounded in the research literature

The mentor can reach the published literature while you talk, so its questions come from the real state of your field rather than from model recall. Finding and reading your own sources stays your work.

Under the hood
OpenAlex, Semantic Scholar, PubMed, arXiv and Europe PMC for papers, ORCID and ROR for people and institutions, Brave or Tavily for the web. The academic ones need no setup, web search needs a free key, and all of them are optional.

It understands what you upload

Papers, documents, data files, photographs of your setup, charts, handwritten notes. The mentor reads what you give it, so it knows what you are actually working with and guides you from there.

Under the hood
Text is pulled out of PDFs by a container running on your own machine, never by an outside service. Reading images and handwriting is a separate option you switch on.

It works through the statistics with you

Upload your measurements and the mentor runs the statistics automatically, then puts the result in front of you pitched to your level and built to teach the reasoning, so the analysis stays yours.

Under the hood
The analysis is automatic rather than something you ask for, and is pitched to your level before any of it reaches the screen. Imported data is capped at 500 MB once decompressed.

It adapts to you

It re-reads where you are on every message, not once at the start: structure where you need it, hints as you find your feet, hard questions where you are strong. Vocabulary matches your age, and it speaks your language.

Under the hood
Each message is classified first, for which part of research it is about, so the tier follows the skill that question actually needs rather than your overall progress. Three tiers by score on that skill: scaffolding below 20, guided discovery to 60, pure Socratic above that. Reflection is a fourth mode rather than the top rung: it is what a strong student gets when the question is about meaning rather than method. Productive struggle is left alone, because that is where the learning happens; unproductive struggle drops you back to scaffolding. Age, cultural background and language are handled at the final presentation step, which translates the whole reply when the language is not English.

It scores eight research skills

Critical reading, data analysis, experimental design, scientific writing, critical thinking, time management, collaboration, and research maturity. Each is scored from the evidence in your own project, so you can see which way it is trending.

Under the hood
Scores run 0 to 100 and each one cites the evidence behind it. Where the evidence is thin, the assessor is told to score low rather than guess. Assessment runs per project and across all your projects, with a trajectory of improving, stable or declining.

A guided path to sharing results

When the work is done: help with the writing, finding the right venue and checking it is not predatory, and taking the work further into talks, media or science art. It is mentorship, not ghostwriting.

Under the hood
One mandatory first step establishes what the research is and who it is for. After that, writing support, venue discovery with predatory-venue detection, and communication guidance are taken in whatever order you want rather than as a fixed sequence. A completion step closes the loop by turning what you found into new questions.

We never see your work, unless you want to

Your conversations and your files stay in a folder on your own machine. No accounts, no telemetry, no server of ours: the Society of Teen Scientists and Authentic Research Partners collect nothing, and we see your work only if you explicitly send it to us. Whichever AI model you run it on does see your conversation.

Under the hood
The software runs locally on your machine, and you bring your own LLM: Claude Code CLI, a local vLLM server, or any OpenAI-compatible API. With Claude Code CLI the model provider is Anthropic; with a remote API it is whoever you configure; with a vLLM server on your own GPU no conversation leaves your computer at all. Searches still reach the databases above, screened for personal details first so your name or your school never lands in a query.
Download and install your own Personal AI Research Mentor

While the software is built to be intuitive, issues may come up. We recommend applying for SoTS membership, which includes email support with a reply within one week.