2. CONFIDENTIAL
Traditional data protection approaches
are ineffective in AI
2
• Static Data Handling. Application, Data,
and Compute are separate
• Fixed, predefined functions
• Rule-based, controlled UX
• Mostly structured data
• Data, Application logic, and compute are
bundled. Continuously learns/evolves
• Adaptive, Dynamic
• Conversational, unpredictable user interactions
• Unstructured/structured,- variety of data, multi-
modal
Vs.
Traditional Applications Gen AI Applications
Standard security controls - Role-
based access, Encryption, etc
Security/controls need to be
defined
3. CONFIDENTIAL
Regulations elevates complexity in enterprise
data protection
WWW.PROTECTO.AI 3
Statute/Bill in
Legislative Process:
Introduced
In Committee
Cross Chamber
Cross Committee
Passed
Signed
Privacy
regulations
are tightening – GDPR,
CCPA, LGPD, POPI,
HIPAA
Canada
Digital Privacy Act
November 2018
California
Consumer Privacy Act
July 2020
Brasil
General Data Protection Law
August 2020
Chile
Proposal Data Protection Law
still in Drafting Stage
US
Competing data privacy bills
issued in Congress
EU
ePrivacy Regulation
still in Drafting Stage
Nigeria
Data Protection Regulation
2019
Uruguay
Law on the Protection of
Personal Data and Habeas Data
January 2019
Argentina
Proposal Data Protection Bill
still in Drafting Stage
South Africa
Protection of Personal
Information Act (POPIA)
still to be determined
Uganda
Data Protection and Privacy Act
still to be determined
Singapore
Personal Data Protection Act
(PDPA) - 2012
China
Personal Information Security
Specification - May 2018
Thailand
Personal Data Protection Act
still to be determined
New Zealand
Privacy Bill 34-2
July 2019
Australia
Privacy Act 1988 and
Amendments- March 2014
India
Personal Data Protection Bill
2018
Kenya
Data Protection Bill
still in Drafting Stage
Stronger AI regulations expected across the globe
AI Regulations
5. CONFIDENTIAL
Protecto protects sensitive data while
preserving utility of the data for AI
W W W . P R O T E C T O . A I
5
Protecto
Data
Transform
Name: John Smith
Phone: 456-876-9345
Customer request …
Name: John Smith
Phone: 456 876 9345
Customer request …
KLOJIOU HNLIHUE
987-923-0234
Gen AI Apps
LLMs
Find Sensitive
Data
Privacy Transform
Machine
Understandable
Synthetic Data
1 2 3
Utility of the
Data
Privacy and
Data Security
Enterprise
Data
6. CONFIDENTIAL
WWW.PROTECTO.AI 6
Makes gen-AI apps privacy-preserving,
compliant, and secure in minutes
3 Ways to Consume Protecto Tokenization
APIs Queue Bulk
Sub-second Performance Updates in minutes For large migrations
(Millions/billions of rows)
No Complex Setup
Enterprise ready – SOC2, Gen AI framework integrations
7. CONFIDENTIAL
Data Protection throughout your AI Lifecycle
W W W . P R O T E C T O . A I 7
Build, Train Tune, RAG Deploy Use
Scan & De-identify
Remove PII, Sensitive Data
Response
Prevent PII Leak, Privacy Controls
Prompts
PII/Data Security scan, DLP Filter
8. CONFIDENTIAL
Case Study 1 – Securing contract review bot
(RAG)
WWW.PROTECTO.AI 8
Goal: Gen AI App based on the Retrieval-Augmented Generation (RAG) that uses historical contracts
as context to create review contracts
Data Protection Challenge: The AI agent could expose confidential data from historic contracts, such
as who wrote the contract to whom
Customer: A Large Telco
9. CONFIDENTIAL
Case Study 2 – Enabling Data Residency of
Sensitive Data
Goal: Leverage OpenAI's capabilities for processing sensitive data, specifically driver history and
criminal records
WWW.PROTECTO.AI 9
Data Protection Challenge: Sending personal data violates data regulations and data residency
requirements
Customer: A Large Consumer Tech
10. CONFIDENTIAL
Protecto vs. Alternatives
WWW.PROTECTO.AI 10
Data Masking for PCI
Previous-Gen Data
Masking
Protecto
Private Cloud
SaaS
SaaS
On-Premises
Private Cloud
SaaS
Deployment Options
Accuracy only on certain
elements
No
Very high accuracy
Identify Sensitive Data Regex based
No
Multiple AI/ML Models, Algo,
Regex, LLMs
+ Heuristic models
No
No
Options to expand custom
identification
No
No
Yes
Multimodal
No
Format only
Format, Type, Length Preserving
Format-Preserving Masking
Yes
No
Yes
Mask Unstructured Data
No
Consistent Pseudonymization
Consistent Pseudonymization,
Anonymization
Consistent Tokens/ Data
Integrity
No
No
Model instructions for higher
comprehension of masked data
Masked Data Comprehension
Encryption-key based
Encryption-key based
Stronger Protection
Random number-based tokens
Security
Reduce
Utility
Loss
Identify
Risks
11. CONFIDENTIAL
Founders with deep data experience
WWW.PROTECTO.AI
11
• Second-time entrepreneur. Cofounded and scaled the previous
startup to $10M ARR
• Microsoft Search & AI, Sun Microsystems, Booz & Co
• MBA from Carnegie Mellon
• MS from LSU. Engineering from CEG, Guindy
Amar
Kanagaraj
Founder & CEO
• 18+ years in Apple
• Expert in data engineering
• Led privacy engineering efforts inside Apple
• Handled petabyte-scale data problems
• Engineering from CEG, Guindy
Baskaran
Alagarsamy
Co-founder & CTO
Team
15+ Engineers (Full Time)
Customers:
Kar Global (KAR), Brookfield,
Belcorp, Nokia
Funding:
Angel Investors (Nov 21)
Head of Android Security, Google
Chief Product Officer, 2nd largest
cybersecurity firm
GM of Incubations, Microsoft
CIOs, CTOs of Large tech companies
Product
Version v0.1 Live (Fall 22)
12. CONFIDENTIAL
GTM - Developer centric, Gen-AI platform
focused
12
Data/ AI Marketplaces Partners
2024 Goals: $1M+ ARR
5 Year: $100M ARR, 1000 Customers
Snowflake
Databricks
Open AI
Huggingface
Developer Focused PLG
Integration with
Langchain,
LLamaIndex,
LangSmith
AWS Bedrock
Community
evangelism
Solution Integrators
1 2 3
13. CONFIDENTIAL
Near Term Roadmap
www.Protecto.ai 13
Secure LLMs
• GPTGuard -
Privacy-Preserving
and Secure ChatGPT
for enterprises
Integrations
• Gen AI frameworks
• Opensource
Frameworks
• AWS Bedrock
• Snowflake Cortex
• Developer templates
Vertical Specific
• Vertical (Health,
Finance) specific
sensitive data
Transformations
• Vertical-specific custom
GPT agents
Multi-Modal
• Sensitive information
Identification across
image/video
• Masking, Blurring /
Pixelation
Privacy Engineering
• Differential Privacy
• Privacy Metrics