The Vision
+The oral examination — the viva — is the most human assessment in academia. It is also the most + broken. A student who studies for three years walks into a room for ten minutes. What happens in + those ten minutes depends not on what they know but on who is sitting across from them.
+TWELVE ends that.
+TWELVE is a kiosk-based AI examination system installed on college-owned computers. It is not a + replacement for education. It is a replacement for the single most biased moment in education — the + oral viva conducted by a human examiner who has opinions, moods, favourites, and blind spots.
+We are not anti-teacher. We are anti-bias.
+ +What TWELVE Does
+-
+
- Identifies the student before the session begins +
- Retrieves their submission history +
- Conducts a structured oral examination through voice and text +
- Evaluates responses against the college's own rubric +
- Produces a final score with a complete, auditable transcript +
Every student. Same quality of questioning. Same standard of evaluation. Score reflects only what + they know.
+ +The Name
+Twelve is a number associated with completeness. Twelve months in a year. Twelve hours on a clock + face. Twelve members on a jury. TWELVE the system is exactly that: a complete, full-cycle assessment + that misses nothing, favours no one, and produces a judgment that stands.
+The Core Problem
+The oral examination system in Indian colleges is broken at three levels simultaneously. Each villain + hurts a different group. All three are enabled by the same root cause: there is no system.
+ +What they do: Assign marks based on favouritism, personal relationships, + attendance records, perceived attitude, or simple dislike.
+Who gets hurt: The sincere student — the one who studied, understood the + material, but was never given space to properly demonstrate knowledge.
+Why devastating: Unlike written exams, a viva leaves no record. No transcript. + No evidence. The student has no way to prove what they said or challenge what they were marked. +
+What they do: Submit AI-generated or copy-pasted work without understanding. + Prepare scripted answers for likely questions. Walk out with full marks because the examiner + didn't drill deep enough.
+Why solvable: A dishonest student cannot answer a question they didn't prepare + for. TWELVE asks questions they cannot prepare for — because they are generated from the + student's own submission in real time.
+What they do: Run oral exams with no consistent rubric, no audit trail, no way + to compare results across sections or years, and no appeals process.
+Who gets hurt: Everyone. Parents who cannot understand marks. Students who + cannot appeal without evidence. Regulators with no visibility.
+The root cause: No one built a system to replace human subjectivity in oral + examination with something better. TWELVE is that system.
+The Solution
+TWELVE replaces the biased human examiner in the oral viva with a structured, AI-driven interview + system. The system:
+-
+
- Authenticates the student before the session begins +
- Reads their actual submitted work before asking a single question +
- Generates personalised, case-based questions from that work +
- Drills down on every answer with cross-questions +
- Detects gaps between claimed understanding and actual understanding +
- Tests core subject knowledge based on curriculum +
- Evaluates every response against the college's own rubric +
- Produces a final score using a panel of AI models +
- Stores a complete transcript for appeal and audit +
No favouritism. No mood. No personal history. No recognition of face or name. Only what the student + knows.
+Core Beliefs
+ +Bias is not personal. It is structural.
+Individual professors are not all bad people. The system gives them no constraints, no + rubric, no record, and no accountability. The bias is the inevitable output of a broken + structure. Fix the structure. Fix the bias.
+A viva should test knowledge, not confidence.
+The student who is nervous and knows everything should not be penalised for their nerves. + The student who is confident and knows nothing should not be rewarded for their + confidence. TWELVE separates the two completely.
+Cross-questioning is the only real test.
+Any student can memorise an answer. A student who actually understands a concept can + answer the same concept asked six different ways. TWELVE asks the same concept six + different ways. Memorisation fails. Understanding passes.
+The appeal process is not a weakness.
+The AI decides. The professor reviews on appeal. This is not a compromise. This is the + correct structure. The AI is the examiner. The professor is the appellate authority. The + transcript is the evidence.
+Transparency is the product.
+Every question TWELVE asked. Every answer the student gave. Every score with reasoning. + All of it stored. All of it accessible. Nothing hidden. Transparency is not a feature. + It is the foundation.
+Start narrow. Prove the core. Then scale.
+TWELVE starts with BTech CSE viva examinations. One subject domain where answers are + objective, verifiable, and defensible. Prove it works there first. Then expand.
+Session Flow
+This is what happens when a student sits down for their TWELVE viva. Every step is deliberate. Every + step serves a purpose.
+ +Kiosk Activation
+The college computer boots into TWELVE kiosk mode. No other application is accessible — + no browser, no file explorer, no task manager. The student sees only the TWELVE + interface.
+Student Identification
+Camera activates. TWELVE prompts the student to state their Name and Roll Number. The + student types or speaks their credentials. TWELVE cross-verifies against the college's + student database. At POC stage, identification is by Name and Roll Number only. Camera + is active exclusively for eye-contact monitoring, not identity verification.
+Submission Retrieval
+Once identity is confirmed, TWELVE silently retrieves the student's submission data: + project submission, Problem Statement assigned, assignment history for the current + semester, and any prior assessment records.
+Pre-Session Analysis
+TWELVE reads the student's submission and builds a question tree silently while a welcome + screen is shown. The AI identifies key technical terms used, the PS and approach taken, + technologies mentioned, and any claims that can be probed.
+PS Verification and Project Deep Dive
+TWELVE opens with the student's own Problem Statement, then drills into the submission + using the student's own words against them. Questions cannot be prepared for because + they are built from the student's submission in real time.
+Cross-Questioning
+The same concept is asked again from a different angle, in a different scenario, with + different framing. A student who memorised an answer will give the same scripted + response to all variants. A student who understood will adapt. TWELVE detects the + difference.
+Core Subject Knowledge
+After project-specific questioning, TWELVE shifts to curriculum-based knowledge for the + student's current semester. Questions are randomised from the curated question bank.
+Session Close and Feedback
+TWELVE closes with brief spoken feedback: what areas were strong, what needs development, + one concrete suggestion for improvement. This feedback is developmental only — not part + of the score.
+Score Generation and Logging
+After the student leaves: final score computed by AI panel, full transcript saved, + session log stored for audit and appeal, next candidate called.
+System Architecture
+The AI Panel
+TWELVE does not use a single AI model to evaluate answers. It uses a panel of multiple models — each + trained on a different dataset with a different reasoning approach — to evaluate every student + response independently.
+ +Evaluates factual and technical correctness against academic standards. Trained on CSE + textbooks, university syllabi, and exam papers.
+Evaluates whether the student can apply the concept. Trained on industry problem-solving + patterns, technical interviews, and case studies.
+Evaluates internal consistency across the session. Trained on cross-examination transcripts + and Q&A sessions.
+The Aggregator
+After all three models score independently, the Aggregator collects scores, applies the college's + pre-defined rubric weightage per section, computes a weighted rational average, and flags any answer + where panel disagreement exceeds a defined threshold for potential human review.
+ +Key principle: Three models trained on different data are not an echo chamber. + They will produce genuinely different evaluations. The aggregator is not averaging sameness — it + is reconciling difference.
+Questioning Strategy
+ +Layer 1 — Submission-Specific Questions
+Generated from reading the student's own submitted work. No two students receive the same questions + because no two students submitted the same work. Questions never ask what the student already wrote + — they ask what follows from it.
+ +Layer 2 — Drill-Down Questions
+For every answer given, TWELVE generates a follow-up that goes one level deeper.
+Layer 3 — Case-Based Variant Questions
+Same concept. Different scenario. Different framing. Asked to verify understanding is transferable. A + student who memorised "merge sort is O(n log n)" will give the same answer to all three variants. A + student who understands sorting will give three different answers.
+ +Layer 4 — Core Subject Questions
+From the curated question bank. Semester and year specific. Randomised per session. Cannot be + predicted or prepared for as a specific set.
+Marking and Evaluation Framework
+ +Who Defines the Rubric
+The college defines the marking rubric before the examination begins. Professors set the criteria + beforehand: what sections are assessed, what weightage each section carries, what constitutes a + passing answer, and what the total marks are. TWELVE applies the institution's standards with + perfect consistency — something a human examiner cannot guarantee across 60 students over two hours. +
+ +What Gets Scored
+-
+
- Project understanding — PS explanation, approach, and implementation reasoning +
- Cross-question performance — ability to handle variant questions +
- Core subject knowledge — curriculum-based questions +
What Is Never Scored
+-
+
- Eye contact — anti-cheat flag only, never a mark +
- Vocal confidence — feedback only, never marks +
- Speed of answering — time taken is logged but not scored +
- Appearance or presentation +
The Appeal Process
+ +The AI decision is the default final mark. On appeal, initiated by the student within a defined + window, the professor reviews:
+-
+
- The full session transcript — every question TWELVE asked, word for word +
- Every answer the student gave, transcribed and timestamped +
- The AI's score for each answer with the reasoning provided +
- The marking rubric used for evaluation +
- Any eye-contact flags logged during the session +
The professor may override the AI's mark if they find a clear and articulable error. The override + must be documented — the professor must state what was wrong with the AI's scoring for the override + to be recorded.
+ +This is not a compromise. This is the correct structure.
+The AI is the examiner. The professor is the appellate authority. The transcript is the evidence. + This mirrors how functional judicial systems work — the trial court decides; appeals go to a + higher authority; the record is the evidence.
+If a student disputes their mark, the professor can point to the exact transcript. There is no "the + professor didn't like me" argument available because the professor did not conduct the examination. + The override itself is also logged and auditable.
+Technology Stack
+ +Frontend — Kiosk Interface
+-
+
- Kiosk shell: Electron.js — cross-platform desktop kiosk, OS-level lockdown +
- UI framework: React with Tailwind CSS +
- Speech-to-text: OpenAI Whisper (local) +
- Text-to-speech: ElevenLabs API with pyttsx3 local fallback +
- Eye-contact analysis: MediaPipe Face Mesh via OpenCV +
Backend — Session Engine
+-
+
- API server: FastAPI (Python) — high performance, async +
- Session management: Redis +
- LLM integration: Anthropic Claude API (primary), OpenAI GPT-4o (fallback) +
- Fine-tuned models: Hugging Face Transformers +
- Document parsing: PyMuPDF, python-docx, pypdf +
AI Panel Models
+-
+
- Model A (Curriculum): Mistral 7B fine-tuned on academic CSE textbooks +
- Model B (Application): LLaMA 3 fine-tuned on technical interview datasets +
- Model C (Coherence): Mistral 7B fine-tuned on Q&A transcripts +
- Question Generator: Claude API (prompt-engineered) +
Database
+-
+
- Primary DB: PostgreSQL with JSONB support +
- Session cache: Redis +
- File storage: Local filesystem (POC) then AWS S3 at scale +
Kiosk OS Lockdown
+-
+
- OS: Windows 10/11 LTSC — most common in Indian college labs +
- Kiosk mode: Windows Assigned Access with Electron +
- USB blocking: Windows Group Policy +
- Network: Local network only during session +
Roadmap
+ +2 weeks
Prove the Core
+Submission parser. Question generator via Claude API. Student answer interface. Basic + scoring. Transcript output. Local prototype only.
+4–8 weeks
First Real Batch
+Kiosk mode. Full session flow. Identity verification. Multi-model panel. Eye-contact + monitoring. Admin portal. Run one real batch at own college in parallel with traditional + viva.
+First External Sale
+Comparative pilot data ready. Approach two to three nearby engineering colleges. + Multi-college isolation. Formal admin portal. Begin question bank expansion.
+Scale
+LMS integration. Multi-department expansion. Self-service onboarding. Biometric verification + post legal clearance. 20+ college target.
+The final test for every feature:
+Does this make the viva fairer — or does it just make it faster? If it just makes it faster, + rebuild it until it makes it fairer. A faster unfair examination is not a product. It is a more + efficient injustice.
+The One Thing That Cannot Be Compromised
+The transcript is sacred. Every question TWELVE asked. Every answer the student gave. Every score + with its reasoning. Stored. Immutable. Accessible.
+If TWELVE ever deletes a transcript, edits a transcript, or hides a transcript — it has become the + same biased examiner it was built to replace. The transcript is the difference between TWELVE and + everything that came before it.
+