Your codebase,
interrogated.

CodeSage analyzes your GitHub repository and generates viva-ready questions that go beyond surface documentation — reaching into architecture, trade-offs, and your actual implementation.

codesage — viva session

$ codesage analyze github.com/you/your-project ✦ Cloning repository...✦ Parsing AST with Tree-sitter (TypeScript, Python)...✦ Embedding 847 code chunks via NV-Embed-QA...✦ Generating examination with Llama-3.3-70B... ✓ Ready · 12 questions across 4 difficulty levels Q1 Your AuthMiddleware validates JWT on every route, but your WebSocket handler skips it. Was this intentional? What are the implications for session hijacking?

Examination Modes

Three ways to be tested

Each mode is calibrated to a different evaluation context, from academic defense to industry hiring loops.

Viva Voce

Oral Defense

Simulates academic examination panels. Questions probe rationale, alternatives considered, and design intent — not just what you built, but why.

Technical Interview

Industry Hiring Loop

Replicates FAANG-style system design and code review rounds. Surfaces scalability gaps, edge cases, and algorithmic trade-offs in your actual codebase.

Code Review

Peer Review Sim

Generates senior-engineer-level feedback on architecture, coupling, test coverage, and documentation framed as questions you should be able to answer.

Specialized NIM models

15+

Languages via Tree-sitter

~3s

Time to first question

Difficulty levels

AI Pipeline

Five models.
One examiner.

No single model handles examination well. CodeSage routes each task to a specialist — embedding, ranking, generation, scoring, and safety run as a coordinated pipeline.

Model	Role	Task
NV-Embed-QA	Embedding	Semantic code chunks
NV-Rerank-QA	Reranking	Query relevance scoring
Llama-3.3-70B	Examiner	Question generation
Nemotron-340B	Scorer	Objective answer grading
Llama-Guard-3	Safety	Content moderation

Live Demo

Questions from real code

These questions were generated from an actual Next.js repository. The examiner read the source — not the README.

src/middleware/auth.tsTS

import { NextRequest, NextResponse } from 'next/server'import { verify } from 'jsonwebtoken'// ← examined by CodeSage export function authMiddleware(  req: NextRequest) {  const token = req.cookies.get('session')  if (!token) return NextResponse.redirect('/login')   try {    verify(token.value, process.env.JWT_SECRET!)    return NextResponse.next()  } catch {    // silent fail — returns undefined  }}

Session score

Correctness

84%

Depth

71%

Security

42%

Architecture

68%

Generated questions

Q 01 · AUTH.TS:10

The catch block on line 14 swallows the error silently. If JWT verification fails, the function returns undefined — not a redirect. How does Next.js handle a middleware that returns undefined?

SecurityMiddlewareHard

Q 02 · AUTH.TS:7

You're reading the token from req.cookies. What are the CSRF implications of cookie-based JWT vs Authorization header?

SecurityMedium

Q 03 · ENV

You use a non-null assertion on process.env.JWT_SECRET!. What happens in production if this variable is unset, and how would you guard against it?

ConfigTypeScriptEasy

Q 04 · ARCHITECTURE

This middleware runs at the edge but verify() from jsonwebtoken is a Node.js library. Why might this cause silent failures in Vercel Edge Runtime?

ArchitectureExpert

Direct GitHub import

Paste any public or private repo URL. CodeSage clones, parses, and analyzes — no local setup required.

AST-level analysis

Tree-sitter parses your code into an abstract syntax tree — questions reference actual line numbers, not summaries.

Personalized study guides

After each session, a scored breakdown maps your weak areas to specific concepts worth reviewing before the real exam.

Progress tracking

Historical session scores tracked across runs so you can see which areas have improved — and which need more work.

Objective scoring

The Nemotron reward model evaluates answers against rubrics — not vibes. Scores are reproducible and explainable.

Built-in content safety

Llama Guard 3 runs alongside the examiner, ensuring every generated question stays within appropriate academic scope.

Know your
own code.

Connect your repository and start your first viva session in under a minute.

Your codebase,
interrogated.

CodeSage analyzes your GitHub repository and generates viva-ready questions that go beyond surface documentation — reaching into architecture, trade-offs, and your actual implementation.

codesage — viva session

Three ways to be tested

Each mode is calibrated to a different evaluation context, from academic defense to industry hiring loops.

Viva Voce

Oral Defense

Simulates academic examination panels. Questions probe rationale, alternatives considered, and design intent — not just what you built, but why.

Technical Interview

Industry Hiring Loop

Replicates FAANG-style system design and code review rounds. Surfaces scalability gaps, edge cases, and algorithmic trade-offs in your actual codebase.

Code Review

Peer Review Sim

Generates senior-engineer-level feedback on architecture, coupling, test coverage, and documentation framed as questions you should be able to answer.

Specialized NIM models

15+

Languages via Tree-sitter

~3s

Time to first question

Difficulty levels

Model

Role

Task

NV-Embed-QA

Embedding

Semantic code chunks

NV-Rerank-QA

Reranking

Query relevance scoring

Llama-3.3-70B

Examiner

Question generation

Nemotron-340B

Scorer

Objective answer grading

Llama-Guard-3

Safety

Content moderation

Questions from real code

These questions were generated from an actual Next.js repository. The examiner read the source — not the README.

src/middleware/auth.tsTS

import { NextRequest, NextResponse } from 'next/server'import { verify } from 'jsonwebtoken'// ← examined by CodeSage export function authMiddleware(  req: NextRequest) {  const token = req.cookies.get('session')  if (!token) return NextResponse.redirect('/login')   try {    verify(token.value, process.env.JWT_SECRET!)    return NextResponse.next()  } catch {    // silent fail — returns undefined  }}

Session score

Correctness

84%

Depth

71%

Security

42%

Architecture

68%

Generated questions

Q 01 · AUTH.TS:10

The catch block on line 14 swallows the error silently. If JWT verification fails, the function returns undefined — not a redirect. How does Next.js handle a middleware that returns undefined?

SecurityMiddlewareHard

Q 02 · AUTH.TS:7

You're reading the token from req.cookies. What are the CSRF implications of cookie-based JWT vs Authorization header?

SecurityMedium

Q 03 · ENV

You use a non-null assertion on process.env.JWT_SECRET!. What happens in production if this variable is unset, and how would you guard against it?

ConfigTypeScriptEasy

Q 04 · ARCHITECTURE

This middleware runs at the edge but verify() from jsonwebtoken is a Node.js library. Why might this cause silent failures in Vercel Edge Runtime?

ArchitectureExpert

Direct GitHub import

Paste any public or private repo URL. CodeSage clones, parses, and analyzes — no local setup required.

AST-level analysis

Tree-sitter parses your code into an abstract syntax tree — questions reference actual line numbers, not summaries.

Personalized study guides

After each session, a scored breakdown maps your weak areas to specific concepts worth reviewing before the real exam.

Progress tracking

Historical session scores tracked across runs so you can see which areas have improved — and which need more work.

Objective scoring

The Nemotron reward model evaluates answers against rubrics — not vibes. Scores are reproducible and explainable.

Built-in content safety

Llama Guard 3 runs alongside the examiner, ensuring every generated question stays within appropriate academic scope.

Your codebase,interrogated.

Three ways to be tested

Five models.One examiner.

Questions from real code

Know yourown code.

Your codebase,interrogated.

Three ways to be tested

Five models.One examiner.

Questions from real code

Know yourown code.

Your codebase,
interrogated.

Five models.
One examiner.

Know your
own code.

Your codebase,
interrogated.

Five models.
One examiner.

Know your
own code.