Privacy Policy

Sarvabhaum AI operates Tensor, an AI-powered academic chatbot for engineering students, accessible at https://sarvabhaum.ai. For data protection purposes, we are the data controller — the entity that determines how and why personal data is processed.

Contact:

• General inquiries: contact@sarvabhaum.ai
• Privacy and data requests: privacy@sarvabhaum.ai
• Address: [Lalitpur, Nepal]

---

2.1 Account Information

When you register, we collect:

• Email address — used for account creation, verification, and communications
• Name — displayed in your profile
• Password — stored as a one-way bcrypt hash (cost factor 12); we never store your plaintext password
• Academic details (optional) — program, batch year
• Profile details (optional) — avatar URL, biography

2.2 Chat and Conversation Data

When you use the chatbot, we collect:

• Your messages and the AI-generated responses
• Metadata about each message: timestamp, number of tokens used, the AI model that generated the response
• Conversation titles and summaries
• Feedback you provide (thumbs up/down, flagging incorrect answers)

2.3 Personal Documents

When you upload study materials, we collect:

• The uploaded file (PDF, image, or document)
• Extracted text content and semantic chunks derived from the file
• Vector embeddings generated from the content (stored in our vector database)
• File metadata: filename, file type, size, and a content hash

Personal documents automatically expire after 7 days. They are purged from our database on your next upload or the next time you access the documents section.

2.4 Study and Learning Data

To power study features, we collect and store:

• Study plans and study schedule entries
• Flashcard content and your spaced-repetition review history
• Topic progress (subjects studied, chapters completed, topics marked as done)
• Past question attempt records (whether you answered correctly, time spent)
• Exam schedule information provided by your institution

2.5 Usage and Performance Data

• Query count and daily token usage (used to enforce free-tier limits and, in anonymised form, for billing audit purposes)
• The AI provider and model used for each response (e.g., DeepSeek, OpenAI)

2.6 Session and Device Data

• IP address — captured in security and authentication logs only; not stored in your profile
• User agent — browser and device type, used for security monitoring
• Device fingerprint — a privacy-preserving hash used to track trial usage for guest (unauthenticated) sessions; not linked to your identity

2.7 Bring Your Own Key (BYOK) API Keys

If you add a personal API key for a third-party AI provider:

• Your key is encrypted using AES-128 (Fernet) before being stored in our database
• The last 4 characters of the key are stored separately so you can identify which key is active — the full key is never shown in the interface
• Keys are only decrypted in server memory at the moment an API call is made on your behalf

---

Purpose	Legal basis (Nepal Individual Privacy Act 2075)
Provide the chatbot and study features	Contractual necessity (you consented at sign-up)
Verify your email address	Contractual necessity
Enforce usage limits and prevent abuse	Legitimate interest
Investigate security incidents and fraud	Legitimate interest
Send transactional emails (verification, password reset)	Contractual necessity
Anonymised analytics to understand usage patterns	Legitimate interest
Improve answer quality using aggregate, anonymised signals	Legitimate interest

We do not use your individual conversations to train AI models. We do not sell your data to any third party.

---

4.1 Tensor is an AI Chatbot

Tensor generates responses using large language models (LLMs). Responses are produced by AI and may be inaccurate, outdated, or incomplete. They should not be relied upon as a substitute for official course materials, instructor guidance, or professional advice.

4.2 Which AI Providers Process Your Messages

When you send a message, it is transmitted to a third-party LLM API to generate a response. The active provider depends on admin configuration. Providers currently supported include:

• DeepSeek (default) — DeepSeek AI
• OpenAI — OpenAI, L.L.C.
• Groq — Groq, Inc.
• Google Gemini — Google LLC
• Mistral AI — Mistral AI SAS

If you have activated a Bring Your Own Key (BYOK) provider, your messages are processed by that provider using your API key, and usage is billed to your account by the provider directly.

Each provider has their own privacy policy. We recommend reviewing the policy of the provider currently active on your account:

• DeepSeek Privacy Policy
• OpenAI Privacy Policy
• Groq Privacy Policy
• Google Privacy Policy
• Mistral AI Privacy Policy

We contractually prohibit providers from using your inputs to train their models where such restrictions are available.

4.3 We Do Not Train on Your Conversations

Your individual conversations are never used to fine-tune or retrain any AI model. We may use aggregate, anonymised patterns (e.g., "questions about thermodynamics are common") to improve the system, but these cannot be traced back to any individual user.

---

Tensor integrates with the following third-party services to operate. Each may process your data as described:

Service	Category	What they receive	Privacy Policy
DeepSeek / OpenAI / Groq / Gemini / Mistral	LLM (chat responses)	Your messages (in transit)	See §4.2
Pinecone (Pinecone Systems Inc.)	Vector database (semantic search)	Text embeddings derived from your documents and conversations	pinecone.io/privacy
Cohere / Voyage AI / Jina AI / NVIDIA	Reranking (improves search results)	Candidate text passages (per admin config)	Respective provider policies
LlamaCloud (LlamaParse)	Document parsing	Your uploaded PDF/document files (admin-uploaded content)	cloud.llamaindex.ai/privacy
Mistral OCR / Unstructured.io	Document parsing	Your personal uploaded documents (per admin config)	Respective provider policies
Resend	Email delivery	Your email address	resend.com/privacy
Cloud storage provider	File storage	Your uploaded documents	Provider-dependent (AWS S3, Cloudflare R2, Google Cloud Storage, Azure Blob, DigitalOcean Spaces, MinIO, or Backblaze B2 — per our deployment configuration)
LangSmith (LangChain, Inc.)	Debug tracing (optional)	Anonymised traces of AI pipeline steps — enabled only by admin	smith.langchain.com/privacy
Umami Analytics	Web analytics	Anonymised page views and navigation events — no personal data, no cookies	Self-hosted on our servers; data never leaves our infrastructure

---

If you add your own API key:

• It is stored encrypted (AES-128 Fernet) in our PostgreSQL database
• Only the last 4 characters are stored in plaintext for display purposes
• The full key is only decrypted in server memory for the duration of an API call
• If you delete the key or your account, it is permanently removed from our database
• You remain solely responsible for the security of your key and any costs incurred through its use

---

7.1 Cookies

We use one cookie:

Name	Type	Purpose	Expiry
tensor_refresh	HttpOnly, Secure, SameSite=Lax	Authentication — stores your refresh token so you stay logged in	7 days

This is a strictly necessary functional cookie. You cannot opt out of it and still use the authenticated service.

We do not use advertising cookies, cross-site tracking cookies, or analytics cookies.

7.2 Local Storage

We store the following in your browser's localStorage (not cookies — not transmitted to our servers):

• Theme preference (dark/light)
• Onboarding completion flag
• UI state (panel open/closed preferences)
• AI disclosure acknowledgement (whether you have seen the first-session AI notice)

None of these contain personal data.

7.3 Analytics

We use Umami Analytics, which is self-hosted on our own servers. Umami is 100% cookieless — it collects anonymised page views and navigation events with no personal identifiers, no cross-site tracking, and no third-party data sharing. No analytics cookies are set.

---

Data	Retention Period
Account profile and settings	Until you request deletion + 30-day recovery window
Chat conversations and messages	Until you delete them, or account deletion
Personal uploaded documents	7 days from upload (auto-expiry)
Encrypted BYOK API keys	Until you delete the key or your account
Refresh tokens	30 days from issue
Security and authentication logs	Up to 90 days
Anonymised usage metrics	Up to 7 years (required for tax and audit purposes)
Infrastructure backups	Up to 30 days after active-system deletion

On account deletion: When you request deletion, your profile and all associated data (conversations, documents, study plans, flashcards, progress, API keys) are scheduled for permanent removal. You have a 30-day recovery window during which you may cancel the deletion by contacting us. After 30 days, data is purged from active systems and removed from infrastructure backups within a further 30 days.

Cloud file storage and vector embeddings: Due to a current infrastructure limitation, files in cloud storage and vector embeddings in our vector database may persist for up to 30 days after you delete a document. We are working to automate immediate cleanup.

Inactive accounts: We plan to implement automatic account deletion for accounts inactive for 13 months, with email warnings at 30 days and 7 days before deletion. This feature is currently in development.

---

Under Nepal's Individual Privacy Act 2075 (2018), you have the following rights:

Right	How to exercise
Access your personal data	Email privacy@sarvabhaum.ai
Correct inaccurate data	Update in Settings, or email us
Delete your account and data	Email privacy@sarvabhaum.ai — in-app deletion is in development
Export a copy of your data	Email privacy@sarvabhaum.ai — in-app export is in development
Opt out of analytics	Use a content blocker; Umami also respects the Do Not Track browser signal
Withdraw consent	You may withdraw your data collection consent at any time, which will require account deletion

We will respond to requests within 30 days. We may request verification of your identity before processing requests.

---

We implement the following technical safeguards:

• Passwords — bcrypt-hashed with a cost factor of 12; never stored in plaintext
• API keys — AES-128 Fernet encryption at rest
• Transport — HTTPS enforced in production; all data encrypted in transit
• Sessions — short-lived JWTs (60-minute access tokens) combined with a 7-day HttpOnly refresh cookie; refresh token rotation with family-reuse detection
• Production guards — JWT secret and encryption key are required environment variables with no insecure defaults in production mode

No security system is perfect. We will notify affected users of any data breach within 72 hours of discovery, as required by applicable law.

---

Tensor is intended for university students and is not directed at users under 16 years of age. If we become aware that we have collected personal data from a child under 16 without verified parental consent, we will delete that account and associated data immediately. If you believe a child's data has been collected, please contact privacy@sarvabhaum.ai.

---

Tensor is operated primarily for users in Nepal. If you access the service from outside Nepal, your data may be transferred to and processed in countries where our third-party providers are located (including the United States and the European Union). We take reasonable contractual steps to ensure appropriate data protection safeguards apply to such transfers.

---

We will notify you of material changes to this Privacy Policy by email at least 30 days before the change takes effect. The updated policy will be posted at https://sarvabhaum.ai/privacy with a revised "Last Updated" date. Your continued use of the service after the effective date constitutes acceptance of the updated policy.

---

For privacy-related requests or questions about this policy:

• Email: privacy@sarvabhaum.ai
• General contact: contact@sarvabhaum.ai
• Address: [Lalitpur, Nepal]

We aim to respond to all inquiries within 30 days.