Socratic Meta-Cognition & Adaptive Learning Engine
The headline
Every other AI hands you a Wikipedia-shaped answer, you say "got it", and that is the end of it. Explic does not accept "got it". It asks you to say the concept back in your own words, and if you cannot, it changes the metaphor and asks again, until you can derive it yourself.
The restatement is a checkpoint, not a courtesy. Agent C validates the learner’s own words into correct, partially correct or wrong, and the mastery gate sits after Agent C and before the knowledge sink: nothing passes unverified, and a failed check flows back into the brain to fill the blind spot.
What you are buying
You are buying the moment where an AI would normally hand over the answer, and Explic holds it back. A 7-level depth gateway sets the baseline context for whatever level the learner is actually at, three Socratic agents then work in series to make them derive the conclusion instead of receiving it, and what passes the mastery check is cleaned and written into a searchable knowledge base. It is not a chatbot. It is a pipeline that produces structured understanding.
What this replaces
- The night before a deadline, someone drops the problem into ChatGPT and comes back with a plan that looks plausible. Everyone nods in the meeting. Three months later you ask why it was designed that way and the room goes quiet, so you hire for it and schedule the work again. The same understanding, bought twice.
- The moment of understanding happens inside one chat window, and it stays there. Close the window and nothing is left — no artefact, no structure, nothing to search. The next person with the same question starts from the beginning, and pays the same tuition again.
- Turning one session into one reusable concept takes a definition, a prompt contract and engineering. Most teams never do that step, because it is engineering work and not conversation work. AI made answers cheap and understanding scarce, and this is the half nobody built.
Architecture
Three stages left to right: the depth gateway that fixes the baseline context, the three Socratic agents in series, and the knowledge sink. The mastery gate sits between the last agent and the sink, and a loop carries structured concepts out of the knowledge base and back into the next session.
Guardrails
- Typed model output. The guide, validation and success states the model returns come back through @ai-sdk/openai under Zod schemas that must match exactly what the front-end state machine requires, which is what keeps the model from drifting into chat.
- Row-level security on the three business tables. The backend is PostgreSQL; profiles, knowledge_bank and chat_sessions all run with RLS and the condition auth.uid() = user_id, so an unauthorised read is refused in the database kernel instead of being filtered in application code.
- Auth at the edge. Under the App Router, middleware.ts validates the Supabase SSR token and completes the permission redirect at the edge before a protected route renders; proxy.ts renews the session in the background, so authentication and session keep-alive are imperceptible.
What you end up holding
- Adaptive immersive web client — the source repository for the fully responsive UI in React 19, TailwindCSS 4 and Framer Motion, deployable on your own infrastructure.
- Semantic database deployment package — Supabase PostgreSQL carrying the profiles, knowledge_bank and chat_sessions tables, the complete RLS isolation set and the triggers.
- Multilingual i18n routing — the locale dispatch mechanism covering English, Chinese, French, Spanish and Japanese.
- Cognition pipeline definition document — the 7 learner_level definitions, the prompt contracts for the three agents and the Zod schema list. You can edit it and add your own dimensions without coming back to us.
- Data and operations handover — the environment variable and key inventory, the deployment procedure, and a knowledge_bank data export.
2.5 to 3 weeks. A planning range, not a contract commitment; the actual schedule follows the signed scope.
Stack
- Orchestration
- Vercel AI SDK 3.0 · Zod (Structured Outputs) · @ai-sdk/openai
- Data
- Supabase Auth · PostgreSQL (Row Level Security + Triggers)
- Presentation
- Next.js 16 (App Router) · React 19 · Tailwind CSS v4 · Radix UI · Framer Motion
The figures, with their arithmetic
- 7 levels of cognitive depth
- An enumerated structural fact from the source: from age-5 bedtime story through undergraduate and expert, to academic, to first principles. Each level carries its own baseline context.
- 3 Socratic agents in series
- A code-level implementation fact: Agent A maps the abstraction onto a concrete metaphor, Agent B asks the structured question instead of answering, Agent C tests the learner by making them restate. Three cores on one cognition pipeline.
- 3 business tables under row-level security
- An implementation fact: profiles, knowledge_bank and chat_sessions, isolated on auth.uid() = user_id, so an unauthorised read is blocked in the database kernel.
- 5 languages routed
- An implementation fact: English, Chinese, French, Spanish and Japanese, with the same cognition pipeline dispatching by locale.
- Knowledge-bank entries: deliberately no figure
- A design capacity, not yet instrumented. The entry count and the de-duplicated concept count are left blank here on purpose and will be reported with the measurement method attached, which is the only way an outcome number from this work is allowed to appear.
The same shape, three other problems
This is the part that matters if the featured use case is not yours: the same architecture pointed at three unrelated situations, none of them a variation on the case above, and none of them requiring a rebuild. These are transfer directions rather than delivered engagements — the system above is what we have run, and the three below are what its shape solves next. We label the difference rather than blur it.
Enterprise SOP training
a compliance manual becomes interactive questioning. A new hire has to restate the procedure to the AI in their own words, and nothing is released until that restatement passes validation.
Digital-health patient education
the terms in a diagnostic report become concrete metaphor — a rusting pipe for a vascular problem — and the patient is asked to say the aftercare logic back, which is how the assistant confirms they actually understood it.
Financial due diligence and private-market decisions
against an unfamiliar sector such as Web3 or AI chips, the tool forces the business logic to be derived from first principles instead of returning a text summary.
Have a system that needs building?
Write down the situation. You get a written, specific response: no calendar, no discovery call, no fifteen minutes of introductions.
Status
Built in-house. This is a tool we built for ourselves to test whether a pipeline can make a model explain knowledge back rather than hand it over — not a client engagement, and no external organisation runs it. It runs in our own stack.
Built in-house. Running in our own stack.