Building the trust layer for AI agents

AI agents fail in production. Logs lie. Retries hide root causes.
I build systems that measure, trace, and fix agent failures at the protocol layer.

Focus: MCP Reliability · AgentOps · Observability · Tooling
Founder: Vouqis · GIant

MCP server reliability auditing via protocol-layer probes.

What it does: Sends 10 probe types across malformed RPC envelopes, timeout boundaries, and schema edge cases to any MCP server. Returns a Trust Score from 0 to 100.

Real result: Exa MCP endpoint scored 92/100. Failed on a malformed JSON-RPC envelope missing id and params. The null-check probes passed.

CI/CD gate:

vouqis audit --url https://your-mcp-server.com --fail-below 80
# Returns non-zero exit code if Trust Score < 80
# Wire into your pipeline. Block bad deployments.

Run it: github.com/Sasisundar2211/Vouqis26

In every production AI system I have built, the majority of failures traced back to the interface layer: malformed tool calls, JSON parsing failures, unhandled retries, and state drift. Not the model. Not the prompt. The protocol.

I build systems that:

• Probe the protocol layer before failures reach production
• Surface exactly which tool call, envelope, or schema edge case broke
• Produce a measurable Trust Score your CI/CD pipeline can gate on

Multi-agent system that detects price drift and enforces contract compliance across vendor documents.

Verified results:

• Standardized extraction outputs across varied document formats
• Reduced manual review cycles per procurement batch

Tech: Python · Gemini 1.5 Flash · NLP · FastAPI · Docker · CI/CD

View Repository

Chains tools with deterministic execution paths. Built as a reliability testbed for tool call validation and retry behavior.

Verified results:

• Stable tool execution across multi-step workflows
• Reusable harness for reliability experiments across agent architectures

Tech: FastAPI · Google ADK · LangChain · Tool execution

View Repository

Before: Manual publishing pipeline, ~2 hours per run After: Automated pipeline, ~10 minutes per run

View Repository

Every system I ship meets this before it is called done:

• Runs on a clean machine
• Dockerized
• CI pipeline validates execution
• Measurable success metric included
• 5-minute install path

Agent Systems

• LLM execution pipelines
• MCP server integration and reliability testing
• Tool orchestration and validation
• RAG systems

Backend

• Python · FastAPI · Function calling systems
• Protocol-layer instrumentation

Infrastructure

• Docker · CI/CD · GCP · AWS
• Observability pipelines

• Built 15+ AI systems with production CI/CD and measurable outputs
• Internship at BITS Pilani Hyderabad
• Nationally selected: Microsoft and SAP TechSaksham program
• Co-authored applied IoT research paper
• Led GenAI prototyping team

• Reliability over features
• Constrained agents over autonomous agents
• Measurable outcomes over claimed improvements
• Kill what does not work rather than ship it

Run Vouqis against your MCP server.

git clone https://github.com/Sasisundar2211/Vouqis26
cd Vouqis26
npm install
vouqis audit --url https://your-mcp-server.com

Get your Trust Score. Find out what your logs are not showing you.

Or keep debugging blindly.

• LinkedIn: linkedin.com/in/sasi-sundar
• Email: sasisundhar2211@gmail.com
• Vouqis: github.com/Sasisundar2211/Vouqis26