Vighnesh Sivaraman
Building production agentic-AI systems and the security around them. ~7 years shipping software across India and the US.
Who I Am
I'm a Senior AI Engineer based in Chennai, India, with ~7 years of software engineering experience.
I build production agentic-AI systems — MCP tooling, agent authentication and LLM automation on AWS and GCP — and the security that lets them pass enterprise reviews and external pen-tests.
I've worked with Claude, Gemini and OpenAI models, MCP servers and multi-agent frameworks. I hold an MS in Computer Science from UT Arlington (thesis track) and co-authored a peer-reviewed paper at ACSAC 2023.
Outside work I build side projects, such as an LLM multi-agent trading-research platform and a land-intelligence MCP server, both running on Google Cloud.
What I Work With
The tools I use to build and secure production agentic-AI systems.
Generative & Agentic AI
AI Security & DevSecOps
Cloud
IaC & DevOps
Languages
Backend & Data
Frontend
ML Foundations
Experience
Sr. Software Engineer / Generative AI Engineer
CurrentKripya Solutions Private Ltd
- ▸ Built zero-standing-credential authentication for a Fortune-10 customer’s autonomous AI agents: STS-verified SigV4 caller identity and cross-account AssumeRole → Secrets Manager (50+ tests), later reused as the shared MCP framework’s SigV4 provider; fixed SigV4 auth for Bedrock agents; wired RBAC and the audit trail into the PLM MCP server and wrote its security-architecture document.
- ▸ Customer point of contact for the auth workstream: presented the MCP authentication approach to the customer’s engineering team (commended by the CTO) and drove the build plan through delivery.
- ▸ Built Jira and Confluence MCP tools for AI agents (issue fetch/upsert/comment, incident summaries, multi-edit Confluence upsert) and the auth, security and load-testing layers of a production PLM MCP server (NestJS, AWS ECS); completed the agent-delegation backend.
- ▸ Fixed a production auth failure on a strategic customer’s two-node tenant (0/20 → 20/20 requests) by isolating the load-balancer stickiness cookie in controlled tests and tracing the root cause to the app-server clustering configuration.
- ▸ Hardened AI platforms for external security testing: made the production Terraform root multi-account for a security-test environment (prod plan unchanged); added AWS WAF rules (allowlist 772 → 5; KnownBadInputs blocking after 741 hostile / 0 legitimate hits) and MFA-gated IAM; fixed a CVSS 8.7 CORS flaw (AppSec-approved); ran an AI-assisted security review of an AI agent platform (controls mapped to the OWASP LLM Top 10) and fixed or mitigated 10 findings (6 rated High) test-first, verified live with two real identities.
- ▸ Delivered the MVP of SCOPE, an ITIL AI co-pilot for global support in Microsoft Teams, and built its Jira/Confluence agent tools and webhooks that generate Claude-written root-cause reports (PDF/DOCX) at ~$0.065 per ticket.
- ▸ Took the lead on most issues in a retail customer’s GenAI design pilot: built an Adobe-style canvas zoom praised in pilot feedback and integrated automatic model selection across 5 image-generation models.
- ▸ Validated, fixed and extended an agent-generated k6 load-testing harness for MCP servers (SSE / Streamable HTTP) and wrote a PKCE-OAuth suite; an ECS autoscaling stress test hit 31,950 / 31,950 requests at p95 293 ms, scaling 2 → 10 tasks, and exposed a health-check grace-period timeout storm.
- ▸ Hardened the supply chain: implemented Google Distroless images (CVEs 194 → 19, 0 Critical/High, SBOM-evidenced; −17.7% size), migrated to pnpm 11 (253 overrides), shipped ~25 tested dependency upgrades; extended CodePipeline/CodeBuild to multiple environments and wrote Terraform for new pre-prod environments.
- ▸ Deployed Kripya’s FinOps product to AWS with Pulumi (VPC, RDS, ECS Fargate, ALB, IAM) and Bitbucket → ECR → ECS CI/CD; split it into three apps with zero resource deletions; closed 81/81 of our department’s findings in an 868-finding AWS security audit.
- ▸ Built the agent-platform layer of Kripya’s Nx monorepo (tenant-aware AgentRegistry, 17 Mastra MCP tools, MCP server, Google ADK migration); co-built a Jira-to-test-case agent (Electron app, SonarQube CI, runtime Gemini / GPT / Qwen switching); designed a 24-hook Claude Code governance catalog and built AI code-acceptance metrics in Grafana.
- ▸ Helped build Kripya’s AI team: wrote the AI Engineer job description and interview assessments, built a screening tool for a 117-candidate campus drive, and mentored an intern.
- ▸ Built a Gemini proof of concept that checks engineering drawings against a 156-point ASME checklist, measured its cost (~INR 10 per model call), and wrote the 10-week project proposal; also compared cloud costs across AWS, Azure and GCP and recommended hardware for running LLMs in-house.
Machine Learning Engineer Intern
Allic Infotech
- ▸ Built a U-Net (ResNet-50 encoder) PyTorch model segmenting skin layers in ultrasound images; shipped it as a Flask web app.
Data Analyst Intern
Revature, LLC
- ▸ Created notebooks for ensuring data warehouse integrity and quality checks in Azure Databricks.
- ▸ Developed visual reports and dashboards using Power BI; gained expertise in data warehouse ETL pipeline design.
Graduate Research Assistant
The University of Texas at Arlington
- ▸ Developed Android applications and ML models for research on "Continuous Authentication using Human-Induced Electric Potential", published at ACSAC 2023.
- ▸ Organized workshops and contributed to a cloud-based IoT system for real-time water quality monitoring.
Senior Software Engineer
Revature, LLC
- ▸ Led a team of five on the Operations & Management team, overseeing daily production bug fixes and feature delivery.
- ▸ Developed microservices for web applications using MySQL, Java Spring Boot framework, and Angular.
- ▸ Managed and improved product infrastructure on AWS and GCP: EC2, RDS, EKS, and Terraform automation scripts.
Projects
Personal and academic projects, from LLM agents and MCP servers to machine-learning fundamentals.
HermesTradingAgent
LLM multi-agent research platform for Indian equities, run in paper-trading mode. A Gemini (Vertex AI) 4-agent debate, episodic memory, post-trade reflection and a weekly LLM critic run on serverless GCP. Every strategy is checked with Deflated Sharpe and Monte Carlo validation, and results are reported honestly, including the ones with no edge. Iterating on its own telemetry (regime-adaptive sizing, trend filters, a backtest gate) cut net loss per INR traded from 0.12% to 0.02% over 101 paper trades: still a loss, and an early signal. 277 commits, 208 tests.
Ellai: Land Intelligence
Streamable-HTTP MCP server on Cloud Run with 13 tools for land due diligence around Chennai: GPS to survey-number resolution from state map data, flood and environmental screening, and Gemini multimodal analysis of Tamil and English title deeds. CI/CD runs a 21-check live smoke test after every deploy.
🚨 Disaster Alert System
Analyzed tweets to classify disaster events using Naive Bayes Classifier with K-fold cross-validation for improved accuracy. Enables real-time disaster detection from social media streams.
🎨 Neura Coloriser
A convolutional autoencoder that colorizes grayscale images. It works in the LAB color space, predicting the color channels from the lightness channel.
Find Route
Implements the A* Search Algorithm for optimal pathfinding between states. Supports both informed (with heuristic) and uninformed search modes using Priority Queues for efficient traversal.
Sentiment Prediction
Predicts the sentiment of text using Naive Bayes classification. Trained and evaluated on labelled datasets to distinguish positive and negative sentiment with measurable accuracy improvements.
🖼️ Titanic — ML from Disaster
Classic Kaggle survival prediction challenge. Applied feature engineering, exploratory data analysis, and multiple classification models to predict passenger survival on the Titanic.
Real-Time Accident Analytics
Published research on a real-time accident detection and analytics system. Built an Android-based data collection pipeline feeding into a web analytics dashboard for incident reporting.
Academic Background
MS Computer Science (Thesis)
The University of Texas at Arlington
Arlington, TX, USA
Thesis: Continuous Authentication using Human-Induced Electric Potential
Relevant Coursework
B.E. Electronics and Communication Engineering
Rajalakshmi Engineering College
Chennai, TN, India
Publications
Continuous Authentication Using Human-Induced Electric Potential
Murali, S., Jin, W., Sivaraman, V., Zhu, H., Ji, T., Li, P., & Li, M.
Proceedings of the 39th Annual Computer Security Applications Conference (ACSAC 2023) · Association for Computing Machinery
Real-time Accident Analytics
Vighnesh Sivaraman et al.
Journal of Chemical and Pharmaceutical Sciences · ISSN: 0974-2115
Certifications
Information Systems Security Professionals
Committee on National Security Systems & National Security Agency
System Administrators Certification
Committee on National Security Systems & National Security Agency
Contact Me
Have a project in mind or want to connect? I'd love to hear from you.