KejiAI

Real statistics. Real interpretation. No signup.

Upload your research data, describe the study, and receive a fully written, APA-formatted academic chapter — calibrated to Nigerian undergraduate, masters, or PhD conventions — in under 60 seconds.

KejiAI is a stateless web app for academic researchers. Students upload a CSV / Excel survey file, fill a short intake form, map columns to objectives and hypotheses, and the system streams a complete chapter back token-by-token with real Pearson r, t-test, ANOVA, chi-square, and Cronbach's α values computed from the uploaded data. Charts are auto-generated as embedded SVGs; the chapter exports as .md, .docx, or PDF (browser print).

No accounts, no database, no payments. The dataset never leaves the request.

---

Quick start

git clone https://github.com/<your-org>/kejiai-chapter-writer.git
cd kejiai-chapter-writer
npm install
# Create .env.local with: ANTHROPIC_API_KEY=sk-ant-…
npm run dev

Open http://localhost:3000.

Verify the Claude connection

node --env-file=.env.local scripts/verify-claude.mjs

---

How it works

User uploads CSV / Excel ──► POST /api/parse-data
                              ├─ papaparse / xlsx
                              ├─ computeStats() — descriptive stats
                              └─ returns { dataset, stats, rawTable }

User maps columns → objectives + hypotheses (in browser)

User clicks Generate ─────► POST /api/generate
                              ├─ rateLimit per IP
                              ├─ runAnalyses() — Cronbach α, Pearson r,
                              │     Welch's t, one-way ANOVA, chi-square
                              ├─ buildCharts() — chart specs as JSON
                              ├─ buildPrompt() — system + user prompts
                              │     with Nigerian academic style rules
                              └─ Anthropic SDK stream → text/plain
                                    ├─ ChapterViewer streams the markdown
                                    ├─ ```chart fences render via Recharts
                                    └─ Tables / headings / source lines
                                       rendered in Nigerian project format

---

Stack

• Framework: Next.js 14 (App Router)
• Language: TypeScript
• Styling: Tailwind CSS + shadcn/ui (editorial cream + amber accent)
• AI: Anthropic Claude API (claude-sonnet-4-20250514) with streaming
• File parsing: xlsx + papaparse
• Charts: recharts (client) + custom server-side spec generator
• Markdown rendering: react-markdown + remark-gfm
• DOCX export: docx
• Rate limiting: in-process token bucket (swap to Upstash for HA)
• Deployment: Vercel (vercel.json pins maxDuration: 90 on /api/generate)

---

What's implemented

• CSV + Excel parsing with numeric / categorical column inference
• Descriptive statistics — mean, SD, min, max, skewness, kurtosis (SPSS conventions)
• Inferential statistics — Pearson r, Welch's t, one-way ANOVA, chi-square, OLS regression, Cronbach's α
• User-driven mapping UI: objectives ↔ Likert items, hypotheses ↔ variable pairs
• Server-side test inference (numeric × numeric → Pearson, cat × num → t / ANOVA, cat × cat → chi-square)
• Streaming chapter generation via Anthropic SDK
• Token-budget-controlled prompts (system ≤ 2000 tokens; total ≤ 4000)
• Nigerian academic style: Decision Rule, "From Table 4.X above", "Source: Field Survey", REJECTED phrasing, Yamane's formula, British spelling, named-stakeholder recommendations
• Auto-generated charts: demographic pies, Likert-item bar charts, scatter for Pearson, group bars for t / ANOVA
• .md, .docx (with embedded chart PNGs), PDF (browser print) export
• Per-IP rate limiting on /api/generate
• Test scripts: test-stats.ts, test-analyses.ts, test-prompt.ts, test-ratelimit.ts

Optional env vars

All of these are off until you set them. Add at Vercel deploy time.

Variable	What it does
WEB_SEARCH_ENABLED	When "true", enables Claude's web_search tool so citations come from real published papers instead of plausible training-time names.
WEB_SEARCH_MAX_USES	Cap on web searches per chapter (default 5).
RATE_LIMIT_GENERATE_PER_HOUR	Per-IP cap on /api/generate (default 10).
UPSTASH_REDIS_REST_URL, UPSTASH_REDIS_REST_TOKEN	Switches the rate limiter to a distributed Redis-backed limiter. Highly recommended for any public deploy — the in-memory fallback only counts within a single serverless instance.
NEXT_PUBLIC_SENTRY_DSN, SENTRY_AUTH_TOKEN	Enables Sentry error tracking. Source maps upload requires the auth token.
NEXT_PUBLIC_PLAUSIBLE_DOMAIN	Enables Plausible analytics. Set to your deployed domain. Tracks file_parsed, generation_started, chapter_copied, chapter_exported.
NEXT_PUBLIC_PLAUSIBLE_SCRIPT	Override the script URL if you self-host Plausible (default https://plausible.io/js/script.js).

What's next

• Vector chart embedding in DOCX is shipped — modern Word uses the embedded SVG; older versions fall back to the PNG.
• Tests for the SVG export round-trip
• OG image + favicon refresh for the deployed domain

---

Project layout

app/
  api/
    parse-data/      CSV/XLSX → dataset + descriptive stats
    generate/        Stream Claude → markdown
    export-docx/     Markdown + chart PNGs → .docx download
  page.tsx           Intake page (client) — Step 01 / 02 / 03 / generate
  generate/          (reserved for future dedicated generate route)
  layout.tsx         Inter + Playfair Display + Geist Mono
  globals.css        Cream palette + @media print

components/
  FileUpload.tsx        Drop zone + parse trigger
  StatsSummary.tsx      N, variables, Likert candidates
  AnalysisMapping.tsx   Map columns → objectives + hypotheses
  IntakeForm.tsx        Title, level, objectives, hypotheses, …
  ChapterViewer.tsx     Streaming markdown viewer + tabs + export buttons
  Chart.tsx             Recharts renderer + parseChartSpec()
  ui/                   shadcn primitives

lib/
  parseData.ts          CSV + XLSX parsers, type inference
  computeStats.ts       All statistical tests (no third-party stats lib)
  runAnalyses.ts        Routes mapped data into the right tests
  buildPrompt.ts        System + user prompts with Nigerian style rules
  buildCharts.ts        Chart specs (deterministic, no Claude tokens)
  exportDocx.ts         Markdown → Word doc with embedded PNGs
  ratelimit.ts          In-process token bucket
  constants.ts          Likert thresholds, APA style, chapter structure, etc.
  utils.ts              cn() helper

types/index.ts          All domain + result types

scripts/
  test-stats.ts         Numerical assertions (Anscombe, SPSS reference values, …)
  test-prompt.ts        Prompt-contract assertions
  test-analyses.ts      Analyses-engine assertions
  test-ratelimit.ts     Rate-limiter assertions
  verify-claude.mjs     One-call live check of the Anthropic SDK

---

Deployment to Vercel

vercel env add ANTHROPIC_API_KEY production
vercel deploy --prod

vercel.json pins maxDuration: 90 on /api/generate so the stream isn't killed mid-chapter on the hobby plan.

For production rate limiting, set:

RATE_LIMIT_GENERATE_PER_HOUR=20    # default: 10

---

License

Private — no license file. All rights reserved.