Learn from the course.
Select a course and documents, retrieve page-linked evidence, then move through explanation, a question, hints and practice. The local source-reader mode remains separate from provider-powered teaching.
← ALL PROJECTS
PROJECT 06 / 07 · Applied AI / Education Technology
Learning, built around the student.
StudyForge AI — Student Learning Workspace
Course evidence. Guided practice. A student’s next step.
WORKING LOCAL PROTOTYPE
01 / THE CONTEXT
The challenge.
Students move between lecture files, practice, job searches and administrative paperwork. StudyForge explores a coherent workspace that keeps course evidence close to learning, while giving jobs and document preparation their own clearly sourced workflows.
02 / MY ROLE
Work with
purpose.
Developing the student-companion product through requirements, API integration, document retrieval, tutoring workflows and testing. It extends supplied StudyForge v2 and earlier academic Corpus Forge processing work, with AI-assisted implementation; it is not presented as an entirely new or exclusively individual codebase.
03 / ENGINEERING NOTES
01
Built around a FastAPI service and a plain HTML/CSS/JavaScript workspace, with SQLite persistence for selected course materials, conversations, activities and progress.
02
Developed page-aware document ingestion, scoped retrieval, provider adapters, citation validation and a persistent tutor-state workflow. Controlled tests exercise explanations, hints, retries, grading and saved practice.
03
Added supporting jobs-provider adapters and French document-preparation workflows. Live provider behavior is separated from offline contract tests and the remaining production roadmap.
04 / HIGH-LEVEL SYSTEM
From idea to structure.
Public-facing conceptual overview — not a disclosure of internal implementation.
STEP / 01Selected course & documents01
STEP / 02FastAPI · retrieval & tutor services02
STEP / 03SQLite · sources, sessions & attempts03
05 / THE STACK
The tools
behind the work.
PythonFastAPISQLiteJavaScriptHTML / CSSPDF / DOCXRetrievalLLM adaptersPytest
06 / MOTION STUDY
See the concept.
An original, interactive learning journey: choose a course, process a sample document, inspect a citation, try an exercise and follow the practice record. Supporting modules remain distinct from the tutor.
STUDYFORGE / LEARNING JOURNEYAUTHORED SAMPLE · 26 SEC
STEP 01 / 07COURSE
A STUDENT’S STARTING POINT
“Help me learn
this course.”
One course. A clear goal.
A place to begin.
01
Start with a learning goal.
A student selects Plant Biology and chooses to understand water transport.
Scripted portfolio illustration · Authored sample material · No files uploaded or external services contacted · Sample progress resets on Restart.
Scripted learning journey with authored sample text, sample progress and fictional supporting workflows. Not a recording of live AI or connected services.
STUDYFORGE AI / LOCAL WORKSPACE / PROJECT CAPTURES
A student’s path through the workspace.
Authentic local application captures, ordered from course context to practice and progress. Select a view, then enlarge it to inspect the interface.
Working local prototype, under active development. This captured version uses local lexical retrieval. Provider configuration does not establish a verified AI tutoring session or live vector-search RAG.
01
AI Study Tutor
Course selection, document context, consent controls and the conversation area. The tutor is ready but no generated answer is shown. The provider is configured, not yet verified, and retrieval is local lexical search.
Private document names redacted; recording overlay hidden.07 / THE PRODUCT VISION
Three modules.
One student’s day.
The tutor is the main learning workflow. Courses, library and progress are supporting views, rather than extra product pillars.
Find a sourced opportunity.
Saved preferences, provider-backed listings, deduplication and original application links. Himalayas and Adzuna adapters exist; coverage, credentials and licensing remain provider-specific. No auto-application.
Prepare with context.
French preparation workflows, dated official-source references, editable checklists and draft correspondence. Optional private text checks stay local. Information support does not establish authenticity or official eligibility.
08 / IMPLEMENTATION & EVIDENCE
Clear about
what works.
The local source has moved beyond the original architecture proposal. Implementation, controlled verification and live acceptance are different levels of evidence.
Implemented and exercised.
FastAPI and SQLite; a plain HTML/CSS/JavaScript frontend; page-aware ingestion; scoped retrieval; persisted tutor states, hints, retries and practice; jobs-provider contracts; administrative checklists and editable drafts.
62 focused offline service tests passed independently during this portfolio integration. Model and jobs responses in these tests are controlled fixtures.
Visible, with boundaries.
Supplied local captures show the student workspace, source-attributed Himalayas listings, practice controls and recorded activity. The tutor view explicitly reports local lexical retrieval and an AI provider configured but not yet verified.
Screenshots are visual evidence, not an independent replay of provider calls. They do not verify an entire tutoring session, answer quality, semantic retrieval or current job availability.
Prototype, with a roadmap.
End-to-end real-provider teaching and grading, live semantic retrieval quality, production accounts and deployment operations remain open. The portfolio animation is a scripted concept, not captured application behavior.
No production deployment, contest completion or independently validated learning outcome is claimed.
09 / ARCHITECTURE, WITHOUT THE HYPE
A working local core.
A deliberate next layer.
FastAPI + SQLite
- Plain JavaScript workspaceHTML/CSS views call the FastAPI endpoints.
- Page-aware evidenceParsers preserve PDF page references. SQLite stores chunks, versioned vectors and source metadata.
- Two retrieval pathsThe offline fallback uses lexical feature hashing. API embedding adapters and linear cosine lookup exist, with controlled semantic fixtures; live semantic quality is unverified.
- Persistent tutor orchestrationThe original tutor-state proposal is now represented in code: conversation state, activities, hidden rubrics, hints and recorded attempts. Fixture tests do not establish live teaching quality.
Evidence before expansion
- Validate real model behaviorNew course PDF → retrieved evidence → cited lesson → follow-up → exercise feedback → saved progress, with provider consent and measurable error handling.
- Evaluate a dedicated vector indexA scalable semantic index remains a proposed evolution of SQLite’s linear search. Embedding quality, deletion and scope isolation need live acceptance.
- Confirm licensed jobs coverageHimalayas/Adzuna adapters are implemented; production terms and live coverage need review. France Travail remains a possible future integration, not a connected provider here.
- Maintain official-source evidenceThe current code has curated French references. A broader, continuously maintained administrative registry, stronger production identity and operational safeguards remain roadmap work.
10 / LIMITS & NEXT STEPS
Build the trust.
Then widen the scope.
This is a local guest-session prototype, with no production account recovery or public multi-user service. OCR, large-scale retrieval, parser isolation, malware scanning and production operations are unfinished.
Jobs may omit hours or work mode. Administrative references are preparation support and need freshness review. No application, email or government filing is sent automatically.
Next: complete real-provider acceptance, evaluate retrieval and grading quality, strengthen production privacy controls, then review licensed sources and deployment readiness.
Source review and focused offline tests · 5 November 2026PROJECT SCOPE
Working local prototype, under active development. Authentic application captures use explicit privacy redactions; the separate animation uses authored sample material and scripted interactions. The captured version uses local lexical retrieval. Live semantic retrieval, complete real-provider tutoring and production readiness are not claimed. No private specifications, student documents or unpublished submission materials are exposed.