Loading
Hi there, I'm Ujwal!
Ujwal Ramachandran
AI Engineer, AI Security Researcher, Data Scientist
- Agentic AI engineer with a security mindset.
- Three years building production AI at Optum.
- Red-teams open LLMs and measures what breaks.
- Main stage speaker at OASec 2026.
About
Data scientist by profession, security researcher by training, AI builder at heart.
At Optum (UnitedHealth Group) I spent three years putting LLM and computer vision tools into production. Two of those projects brought $5M and $3M in annual savings.
Now I am at NTU Singapore for an MSc in Cyber Security, red-teaming LLMs and researching the AI supply chain. In September 2026 I presented that research, Ghost in the Template, on the main stage at OASec.
Impact
-
Speaker @ OASec 2026
one talk on the single main stage, selected from 60+ submissions
Ghost in the Template
Presented original research on chat-template backdoors in GGUF models at a practitioner AI security conference in Singapore.
-
$3M
a year saved, 70% faster contract reviews
Provider Contracts
Automated document digitization and cross-referencing against multiple sources of truth, with a Confidence Score Metric that routes uncertain cases to reviewers.
-
$5M
a year saved, 45% less manual processing
Contract Redlining Tool
As project lead, built a contract redlining and document comparison tool with OpenAI API LLM summarizers behind Flask REST APIs and a custom GUI for business users.
-
1000s
enterprise contracts processed
Extraction Pipeline
Azure Document Intelligence combined with the team's patented Table Sense table-detection model.
-
90%
fewer vulnerabilities in a regulated HIPAA / PHI environment
Security Gate
Owned the security gate for the Contract Intelligence team's deployments: Docker and CI/CD configuration, Azure Container Apps with Key Vault, Prisma and Azure Defender scanning.
-
10%
sales growth
Price Intelligence
Python and Docker web scraping framework for competitive price tracking with CAPTCHA handling.
-
78%
audited attack success across 450 evaluations
AutoRedTeam-LLM
Model-agnostic red-teaming pipeline, 150 adversarial prompts against three open models. The automated judge was checked against a 45-sample human gold set, then every VULNERABLE label was re-audited and 15 were corrected before reporting.
-
3 regions
PoC demos and requirement walkthroughs for US, Dublin and India teams
Stakeholder Delivery
US Operations, Retail and Pharmacy teams, and teams in Dublin and India, across multiple time zones, turned into delivered AI features.
Work and writing
-
Chat-Template Backdoors in GGUF Models
-
Automated security benchmark for open-source LLMs
-
A personal anti-scam analyst built from specialised agents
-
Automated PCAP analysis at scale
The journey
From first compile to the model supply chain.
Toolkit
- Multi-Agent Systems
- LLMs
- RAG
- Vector Search
- Prompt Engineering
- LLM Red-Teaming
- Safety Evaluation
- Hallucination Mitigation
- Tool Use
- HuggingFace
- sentence-transformers
- 4-bit NF4 Quantization
- Streamlit
- Adversarial ML
- Deep Learning
- Explainable AI
- MLOps
- Reinforcement Learning
- Computer Vision
- YOLO v5
- OpenCV
- PyTorch
- AI Supply Chain Security
- Model Integrity and Provenance
- Zero Trust Architecture
- SIEM / SOAR
- Application Security
- Cryptography
- IAM
- CI/CD Security
- Vulnerability Management
- Secure Coding
- GDPR / PDPA
- HIPAA / PHI
- Python
- Pandas
- NumPy
- Scikit-learn
- Flask
- FastAPI
- REST APIs
- OpenAI API
- SQL
- Shell
- C
- C++
- HTML / CSS
- Azure
- Docker
- Kubernetes
- Jenkins
- Git
- Linux
- MongoDB Atlas
- Oracle DB
- GraphDB
- Hive
- Selenium
- n8n
01 Agentic AI and LLMs
- Multi-Agent Systems
- LLMs
- RAG
- Vector Search
- Prompt Engineering
- LLM Red-Teaming
- Safety Evaluation
- Hallucination Mitigation
- Tool Use
- HuggingFace
- sentence-transformers
- 4-bit NF4 Quantization
- Streamlit
02 AI and Machine Learning
- Adversarial ML
- Deep Learning
- Explainable AI (XAI)
- MLOps
- Reinforcement Learning
- Computer Vision
- YOLO v5
- OpenCV
- PyTorch
03 Security
- AI Supply Chain Security
- Model Integrity and Provenance
- Zero Trust Architecture
- SIEM / SOAR
- Application Security
- Cryptography
- IAM (Keycloak)
- CI/CD Security (Prisma, Azure Defender)
- Vulnerability Management
- Secure Coding
- GDPR / PDPA
- HIPAA / PHI
04 Programming
- Python
- Pandas
- NumPy
- Scikit-learn
- Flask
- FastAPI
- REST APIs
- OpenAI API
- SQL
- Shell
- C
- C++
- HTML / CSS
05 Cloud and Tools
- Azure (AKS, Cosmos DB, Document Intelligence, Container Apps, Key Vault)
- Docker
- Kubernetes
- Jenkins
- Git
- Linux
- MongoDB Atlas
- Oracle DB
- GraphDB
- Hive
- Selenium
- n8n
- Microsoft Certified, Azure Fundamentals (AZ-900)
Driven by curiosity, powered by code, secured by design.
Shell
Talk to the site. Ask the terminal anything it knows. Start with help.
ujwal@eclipse
name: Ujwal Ramachandran
roles: AI Engineer, AI Security Researcher, Data Scientist
based: Singapore
now: MSc Cyber Security, Nanyang Technological University
Type help, or pick a command above.
Off the clock
- AI / ML Intelligent systems, automation, and what LLMs and agents can do when they are built carefully.
- Cybersecurity Threat analysis, defense strategies, and building security architectures that hold up.
- Astronomy Space exploration and celestial phenomena, the things that keep scientific curiosity alive.
- Photography Wildlife and nature, mostly up close. Not professional, just patient.
- Volunteer for environmental and sustainability initiatives.
- Games of every kind, outdoor sports and online.