How I Used Claude to Finish Building an AI That Evaluates AI — and Caught It Hallucinating

vijaya kumari — Sat, 23 May 2026 02:47:43 +0000

The Project I Started But Never Finished

Earlier this year I started building ai-qe-agent —
a multi-agent system that auto-generates QA test cases
using Claude (Anthropic's AI).

8 specialized agents. TypeScript. Direct Anthropic SDK.

It worked. But it had a critical problem:

No visibility into whether the outputs were actually correct.

Agents were generating test cases, reviewing them,
converting them to Playwright scripts — and I had
no idea if Claude was hallucinating, truncating,
or silently failing between agents.

That's what I set out to finish.

The Before

How Claude Helped Me Finish It

I used Claude (via Claude Code) as my primary
AI coding assistant throughout this project.

Claude helped me:

Design the LLM-as-Judge eval architecture
Generate eval_suite.py from scratch
Debug LangSmith tracing integration
Build the TruLens monitoring setup
Create the Fintech AI Agent Gradio app

The meta-irony: I used Claude to build a system
that evaluates Claude's own outputs.

What I Finished

1. Custom LLM Eval Suite

Built eval_suite.py using LLM-as-Judge pattern —
Claude evaluating Claude's own outputs across 4 dimensions:

Completeness — did the agent complete the full task?
Specificity — were outputs precise and detailed?
Faithfulness — did the agent follow all instructions?
Hallucination detection — did it invent facts not in context?

2. TruLens Monitoring Dashboard

Real-time quality metrics across all 4 agents:

Faithfulness scores
Hallucination flags
Chain compatibility checks
Quality score trends

3. LangSmith Production Tracing

Every Claude API call now traced:

Input prompt
Output response
Latency per agent
Token usage

4. Pinecone Vector Store

Semantic deduplication for test cases:

Prevents duplicate test generation
0.85+ cosine similarity = HIGH OVERLAP flag

5. Fintech AI Agent (New HF Space)

Live demo combining everything:

Fraud detection with risk scoring (0-10)
Compliance Q&A (KYC/AML/GDPR/SOX/PCI-DSS)
AML risk report generation (6-section formal reports)
Real-time eval dashboard

The Findings — What Claude Found About Itself

Running the eval suite on my own pipeline revealed:

🔴 2 hallucinations caught

AutomationScriptGenerator invented 'Invalid credentials'
as error text — never specified in the input context.
SelfHealingAgent fabricated DOM selectors without a DOM.

🔴 2 pipeline breaks found

ManualTestGenerator output = bare array.
QAReviewAgent expected a wrapped ManualTestSuite object.
chain_compatibility = 0. Would silently fail in production.

🔴 2 faithfulness failures

ManualTestGenerator generated 2 of 8 required test cases.
Stopped with no error. No warning. Just silent truncation.

🟢 0.902 avg quality score

AutomationScriptGenerator: 0.94
SelfHealingAgent: 1.0 quality — but 0.0 faithfulness.
Good output. Wrong process. Only eval catches this.

The After

The Key Insight

AI systems fail silently.
No errors. No warnings. No crashes.
Just wrong outputs — shipped with confidence.

This is why LLM Evaluation Engineering exists.
And why finishing this project mattered.

Demo

🤗 Fintech AI Agent (live):
https://huggingface.co/spaces/Vijayarv07/fintech-ai-agent

🤗 ai-qe-agent (live):
https://huggingface.co/spaces/Vijayarv07/ai-qe-agent

⭐ GitHub:
https://github.com/vijayarjun7/ai-qe-agent

Tech Stack

Claude (claude-sonnet-4-20250514) — Anthropic
Python + TypeScript
TruLens (eval monitoring)
LangSmith (production tracing)
Pinecone (vector store)
Gradio (HF Space UI)
Playwright (automation)

Built in public. Follow my journey: #BuildInPublic

How I Built a Fintech AI Agent That Detects Fraud and Generates AML Risk Reports — With Zero Hallucinations

vijaya kumari — Fri, 22 May 2026 01:34:28 +0000

The Problem

AI in fintech cannot hallucinate.

A fabricated regulation reference = legal liability.
A missed fraud pattern = financial crime.
A wrong compliance answer = regulatory penalty.

Yet most AI demos ship without any eval layer.
I built one.

What I Built

A 4-tab Fintech AI Agent deployed on HuggingFace:

Tab 1 — Fraud Detector
Analyzes transactions for fraud patterns.
Returns: risk score (0-10), red flags,
approve/review/reject recommendation.

Test input:
"Transfer $9,800 to Cayman Islands
at 3:47am from unrecognized device"

Result: 9/10 HIGH RISK → REJECT

Red flags caught:

Amount just below $10K CTR threshold (structuring)
High-risk jurisdiction (Cayman Islands)
Unusual transaction time (3:47am)
New unrecognized device

Tab 2 — Compliance Q&A
RAG over hardcoded financial regulations:
KYC, AML, GDPR, SOX, PCI-DSS

Every answer:

Cites specific regulation + section
Shows confidence score
Flags hallucination risk

Tab 3 — AML Risk Report Generator
Generates formal 6-section risk assessments:

Customer Risk Profile
Transaction Pattern Analysis
Red Flags Identified
Regulatory Considerations
Recommended Actions
Compliance Officer Notes

Tab 4 — Eval Dashboard
Real-time metrics across all tabs:

Total queries processed
Avg quality score
Hallucinations flagged
Risk alerts triggered

The Eval Layer

Every Claude output is scored for:

{
  "faithfulness_score": 0.95,
  "confidence": 0.85,
  "hallucination_risk": "LOW"
}

This is the LLM-as-Judge pattern —
Claude evaluating Claude's own outputs.

Results from first run:

98% avg quality score
0 hallucinations detected
Faithfulness: 95-100% per tab

Tech Stack

Claude (claude-sonnet-4-20250514)
Pinecone (vector store + semantic dedup)
LangSmith (production tracing)
TruLens (eval monitoring dashboard)
Gradio (HF Space UI)

Live Demo

huggingface.co/spaces/Vijayarv07/fintech-ai-agent

GitHub:
github.com/vijayarjun7

What's Next

Adding quantum-inspired compression to the
inference layer (QuantRot-PQC research).

Because reliable AI + efficient AI =
production-ready AI.

Forem: vijaya kumari

How I Used Claude to Finish Building an AI That Evaluates AI — and Caught It Hallucinating

The Project I Started But Never Finished

The Before

How Claude Helped Me Finish It

What I Finished

1. Custom LLM Eval Suite

2. TruLens Monitoring Dashboard

3. LangSmith Production Tracing

4. Pinecone Vector Store

5. Fintech AI Agent (New HF Space)

The Findings — What Claude Found About Itself

The After

The Key Insight

Demo

Tech Stack

How I Built a Fintech AI Agent That Detects Fraud and Generates AML Risk Reports — With Zero Hallucinations

The Problem

What I Built

The Eval Layer

Tech Stack

Live Demo

What's Next

BuildInPublic