CampusComply : Privacy Preserving AI for Trusted Campus Regulations

Tagline: Understand every rule. Trace every answer. Protect every document.

1. Problem Statement

Campus regulations are fragmented across academic handbooks, examination policies, hostel guidelines, circulars and departmental notices. Students often struggle to locate the exact rule, identify applicable exceptions or determine which provision applies to their situation.

Generic AI assistants can generate plausible but unsupported answers, overlook conflicting documents and expose sensitive information when documents are processed through external AI services.

The problem is not the lack of information. It is the lack of a reliable, traceable and privacy-conscious way to understand it.

CampusComply addresses the gap between having institutional rules and being able to retrieve, interpret and verify them confidently.

2. Solution

CampusComply is an evidence-driven Campus Regulations Intelligence Platform that transforms institutional documents into a searchable, interconnected and auditable knowledge system.

It combines Retrieval-Augmented Generation (RAG), adaptive Gemma model routing, privacy-preserving document redaction, knowledge-graph reasoning and policy guardrails to answer questions using relevant institutional evidence.

Every answer is designed to include source-document citations, PDF page references and supporting passages. When evidence is insufficient, the platform abstains. When documents appear to conflict, it surfaces both provisions instead of inventing a resolution.

From scattered documents to verified answers. From isolated passages to connected regulatory knowledge.

3. Innovative Features

A. Adaptive Gemma Model Router

CampusComply uses four configurable Gemma model candidates spanning Gemma 2, Gemma 3 and Gemma 4. Its application-level router estimates task complexity, input-token requirements, retrieved evidence size and response requirements to select an eligible model.

Simple factual questions can use a lightweight model, while multi-document synthesis and complicated regulatory conflicts can be routed to stronger models when benchmarks justify the additional cost and latency. OpenRouter provides the common inference interface.

B. PRAG — Privacy-Preserving Retrieval-Augmented Generation

Before documents are sent to external AI services, locally processed sensitive regions are replaced with consistent placeholders such as [PERSON_001] or [STUDENT_ID_001].

The sanitized representation is used for embeddings, retrieval, graph construction, classification and generation. Essential regulatory information is preserved wherever possible, while configured high-risk processing failures block downstream inference.

This reduces sensitive-data exposure; it does not claim perfect detection or eliminate the need for privacy testing and review.

C. Regulatory Knowledge Graph

CampusComply converts extracted regulatory knowledge into a directed graph of documents, rules, obligations, deadlines, thresholds, exceptions and affected groups.

Relationships such as REQUIRES, APPLIES_TO, EXCEPTION_TO and SUPERSEDES help identify related provisions across documents. Graph traversal narrows the evidence retrieved for a question, allowing Gemma to focus on relevant rules rather than processing unrelated passages from entire handbooks.

Every extracted relationship retains source provenance so it can be traced to its originating document and page.

D. Jev-Assisted Classification

A structured classification layer using Jev categorizes requests, identifies task types and supports routing decisions where useful. Gemma then performs the natural-language explanation and evidence-grounded answer generation.

The classification component and generation component have separate responsibilities. Jev’s contribution will be evaluated against a deterministic baseline before being treated as a production improvement.

E. Regulatory Conflict Intelligence

CampusComply compares provisions that appear to address the same subject and detects potential conflicts involving thresholds, deadlines, scope, exceptions or document versions.

It displays both supporting passages and their citations. It does not arbitrarily select the authoritative rule when precedence cannot be established from verified source information.

F. Evidence-Gated Answers

A relevance gate checks whether retrieved evidence is sufficient before normal answer generation. Citation validation checks that references point to real retrieved passages, and additional evidence checks assess whether those passages support the associated claims.

When the documents do not provide sufficient evidence, CampusComply abstains rather than filling gaps with general model knowledge.

G. Policy-Aware Guardrails

The platform declines requests seeking actionable assistance to exploit loopholes, bypass legitimate rules or engineer unfair advantages from regulations. It distinguishes these requests from legitimate questions about interpreting rules, understanding exceptions, filing appeals or improving unclear policies.

Guardrails apply before generation and during response validation, with tests for prompt injection, unsafe requests and false refusals.

4. Key Differentiators

1. Privacy before inference. Sensitive regions are scrubbed before external models receive document content; privacy is enforced throughout the processing pipeline rather than only in the final prompt.

2. Graph-augmented retrieval. The system connects related obligations and exceptions across documents, helping it answer questions that require tracing more than one isolated passage.

3. Adaptive intelligence. Model selection is based on task and context requirements, subject to actual availability and measured quality, rather than routing every question to the largest model.

4. Classification separated from generation. Jev can support structured decisions, while Gemma is responsible for natural-language responses. Each component is tested for its own role.

5. Conflict-aware answers. The system can expose apparent contradictions, preserve both sources and communicate uncertainty rather than silently choosing a rule.

6. Measurable reliability. RAG versus no-RAG, retrieval configurations, model variants, refusal behavior, citations, latency and workload cost are evaluated through reproducible experiments.

Our differentiator: CampusComply does not simply retrieve information and generate an answer. It combines privacy controls, connected evidence, adaptive inference and explicit uncertainty handling into a single regulatory intelligence workflow.

5. Best Use of Gemma

Gemma is the primary natural-language reasoning and generation family within CampusComply.

The initial model registry includes four candidates:

• Gemma 2 9B: lightweight baseline for suitable short, factual questions.
• Gemma 3 12B: intermediate candidate for ordinary explanations and modest synthesis tasks.
• Gemma 4 26B A4B: candidate for more demanding multi-source analysis and structured reasoning.
• Gemma 4 31B: candidate for difficult synthesis and complex regulatory questions.

The exact model IDs, provider availability, supported context limits and pricing will be verified through the live model registry. The router will reserve output capacity and reject routes that cannot accommodate the complete prompt.

Gemma is used for:

• Evidence-grounded answer generation.
• Summarizing relevant regulatory passages.
• Explaining how verified provisions relate to one another.
• Producing structured explanations of potential conflicts.
• Interpreting retrieved context while following strict citation and abstention requirements.

Optimization objective: maximize answer correctness and evidence support while controlling latency and total workload cost. The winning model-routing policy will be determined experimentally rather than assumed in advance.

6. Open-Source Approach

CampusComply is designed around modular, inspectable and reproducible components.

• Open application code: publish the RAG core, document schemas, retrieval pipeline, graph design, evaluation harness and integration interfaces under a clearly specified open-source license.
• Open-weight models: use compatible Gemma models with model-specific license verification. Gemma 4 was announced under Apache 2.0 in April 2026; licenses for other generations and variants must be checked independently.
• Interchangeable providers: keep the model interface separate from the core RAG logic so inference providers can be replaced without rewriting retrieval and document processing.
• Reproducible benchmarks: publish configurations, evaluation methodology, non-sensitive test fixtures and measured results.
• Privacy-conscious deployment: keep original documents, sensitive mappings and protected storage under institutional control wherever deployment requirements permit.

Transparency commitment: CampusComply will distinguish open-source application code, open-weight models and hosted third-party services. Using OpenRouter and Jev APIs means those inference services are not automatically self-hosted or offline. A local-inference deployment can be explored separately.

7. System Architecture

Layer 1 — Document Ingestion and Privacy

PDF upload → validation → local text extraction → sensitive-region detection → redaction → sanitized canonical document.

Layer 2 — Knowledge Preparation

Sanitized text → page-aware chunking → embeddings → keyword index → regulatory entity and relationship extraction → knowledge graph with source provenance.

Layer 3 — Request Understanding and Safety

Student question → input checks → guardrail policy → optional Jev classification → query interpretation → complexity and context-budget assessment.

Layer 4 — Evidence Retrieval

Semantic search + keyword search + graph traversal → rank fusion → relevant evidence selection → relevance threshold → cited passage set.

Layer 5 — Adaptive Generation

Context-aware application router → selected Gemma model through OpenRouter → evidence-grounded generation → citation validation → output policy checks.

Layer 6 — Application Experience

CampusComply web interface → answer and source panel → document-page references → potential conflict comparison → conversation history → measured evaluation dashboard.

The core architectural principle is that every component operates on controlled, provenance-preserving data, and no downstream model can bypass a previous safety or privacy decision.

8. Technology Stack

Frontend

• Next.js App Router
• React and TypeScript
• Tailwind CSS
• shadcn/ui and Lucide
• Accessible document viewers, evidence panels and conflict-comparison layouts

Backend

• Python
• FastAPI
• Pydantic for structured schemas
• Shared, modular Python RAG core
• Asynchronous API handling where appropriate

AI and Model Access

• Google Gemma 2, Gemma 3 and Gemma 4 candidates
• OpenRouter API for unified hosted inference
• TypeSafe Jev API for structured classification, if validated by experiments
• Sentence Transformers for embeddings
• BM25 and semantic retrieval with rank fusion

Privacy and Guardrails

• Local PII detection and deterministic redaction rules
• Configurable privacy policy
• Input classification and safety checks
• Evidence-based abstention
• Citation and output validation

Knowledge and Storage

• Neo4j for the directed regulatory knowledge graph
• A persistent vector index for embeddings
• PostgreSQL for document metadata, conversations and audit records
• Protected file storage for uploads and sanitized derivatives

Evaluation and Operations

• Pytest for backend tests
• Automated retrieval and answer-quality benchmarks
• Versioned evaluation datasets
• Structured latency, error and model-usage metrics
• Docker and Docker Compose for reproducible local development

Storage products and service choices remain implementation decisions until the repository is inspected and their operational costs are tested.

9. Why in 2026?

AI capability has advanced, but trust remains an engineering problem.

In 2026, open-weight Gemma models offer a broader range of reasoning, model-size and deployment options. This makes it practical to explore model routing, local preprocessing and model selection instead of treating every question as an identical request to one large model.

At the same time, institutional documents can contain personal information, and Indian data-protection governance is becoming more concrete: the Government of India notified the Digital Personal Data Protection Rules, 2025 on 14 November 2025. CampusComply is designed with privacy-aware processing and controlled data flows in mind, without claiming that the product itself guarantees legal compliance.

Three trends make the project timely:

• More capable open models: enable adaptable AI architectures with different inference trade-offs.
• Greater need for accountable AI: citation quality, abstention, traceability and uncertainty handling must be designed explicitly.
• Growing importance of data stewardship: document-processing systems need stronger privacy controls and clear boundaries around third-party services.

CampusComply applies these trends to a concrete institutional problem where incorrect answers can confuse students and where source traceability matters.

10. End Goal

The ultimate goal is to create a dependable regulatory intelligence platform that transforms institutional knowledge into an accessible, verifiable service for students and administrators.

Students should be able to ask natural-language questions, receive answers grounded in official documents, inspect the exact evidence, understand exceptions and see when policies appear inconsistent. Institutions should be able to update documents, audit retrieval and answer behavior, measure model performance and maintain greater control over the information processed by AI.

The initial release targets campus regulations. The architecture may later support other institutional knowledge domains, including academic handbooks, scholarship rules, hostel policies and administrative procedures, subject to domain-specific validation and access controls.

CampusComply’s end goal is not to replace institutional authority. It is to make institutional rules easier to find, understand, verify and improve—without asking users to trust an answer they cannot trace back to evidence.