{
    "service": "Tachy AI Education Agent",
    "domain": "ai.tachy.in",
    "version": "0.1.0",
    "manifest": {
        "vision": "A dedicated education AI agent for handwritten answer checking, question generation, online exams, and student performance conversations.",
        "principles": [
            "Keep ai.tachy.in isolated from existing production projects.",
            "Build a secure education workflow before enabling public uploads.",
            "Use model orchestration and evaluation instead of claiming real AGI before proof.",
            "Keep every student document auditable, permissioned, and removable."
        ],
        "first_modules": [
            {
                "name": "Document intake",
                "status": "design-ready",
                "scope": "PDF/image upload, institution metadata, exam metadata, consent, retention rules."
            },
            {
                "name": "Handwriting OCR and layout understanding",
                "status": "research",
                "scope": "Page splitting, answer-region detection, handwriting OCR, diagrams, tables, and math handling."
            },
            {
                "name": "Answer key and copy checking",
                "status": "research",
                "scope": "Rubric matching, semantic similarity, plagiarism/copy-pattern signals, and human review queue."
            },
            {
                "name": "Question generation",
                "status": "research",
                "scope": "MCQ, short answer, long answer, true/false, fill blanks, case studies, and viva prompts."
            },
            {
                "name": "Online exam engine",
                "status": "planned",
                "scope": "Question bank, exam scheduling, proctoring hooks, submissions, scoring, analytics."
            },
            {
                "name": "Student performance agent",
                "status": "planned",
                "scope": "Student chat, improvement plan, weak topic detection, teacher-visible audit trail."
            }
        ],
        "seven_day_plan": {
            "day_1": "Finalize product scope, roles, privacy rules, data retention, upload formats, and first institution workflow.",
            "day_2": "Define architecture, database schema, document pipeline, queue system, and model-provider decision matrix.",
            "day_3": "Prototype OCR pipeline on sample handwritten copies and measure extraction accuracy by subject and language.",
            "day_4": "Prototype answer checking with rubrics, model scoring, copy-check signals, and manual review overrides.",
            "day_5": "Prototype question generation from scanned material and validate MCQ/short/long-answer quality with teachers.",
            "day_6": "Design student/teacher/admin UX, permission model, audit logs, and reporting dashboards.",
            "day_7": "Freeze MVP backlog, cost estimate, model stack, rollout phases, and security checklist before full build."
        },
        "model_strategy": {
            "phase_1": "Use strong hosted vision-language and text models for OCR assist, reasoning, rubric scoring, and chat.",
            "phase_2": "Add retrieval over school content, syllabus, textbooks, solved papers, and institution policies.",
            "phase_3": "Fine-tune or adapt smaller models only after collecting permissioned, labeled education data.",
            "phase_4": "Build a multi-agent layer: document agent, evaluator agent, question-maker agent, tutor agent, and audit agent.",
            "agi_note": "The project can be AGI-inspired and expandable, but production claims must stay evidence-based and measurable."
        },
        "risk_controls": [
            "Uploads stay disabled until authentication, rate limits, malware scanning, and storage retention are configured.",
            "AI scoring must include confidence, evidence, and teacher override for academic decisions.",
            "Student data must be encrypted at rest, permissioned by institution, and deleted by policy.",
            "Medical or high-stakes content must be educational support only unless reviewed by qualified experts."
        ]
    }
}