Intelligent Log Analyzer

Guardrailed AI-Assisted Incident Intelligence Service

A hybrid deterministic + LLM-powered incident analysis system designed to improve incident triage reliability while reducing hallucination risk in production environments.

---

System Overview

This system combines:

• Deterministic log parsing
• Structured metrics extraction
• Constrained LLM reasoning
• Strict schema validation (Pydantic)

The goal is to balance reliability with contextual intelligence.

---

Architecture

---

LLM Guardrails

• JSON-only output instruction
• Pydantic schema validation before rendering
• Low temperature (0.2) for reduced variability
• Controlled log truncation to manage context size
• Output rejection on validation failure

---

Example Scenario 1 --- Database Connection Pool Exhaustion

Input Logs

2024-01-15 09:13:22 ERROR Failed to acquire database connection from pool
2024-01-15 09:13:24 ERROR Timeout while waiting for database connection
2024-01-15 09:13:50 ERROR Service degraded: high response latency detected

Output Screenshot

---

Example Scenario 2 --- Memory Leak / OOM Restart

Input Logs

2024-03-02 14:05:15 WARN High memory usage detected: 85%
2024-03-02 14:15:44 WARN High memory usage detected: 92%
2024-03-02 14:18:30 ERROR Process killed due to OutOfMemory
2024-03-02 14:18:31 ERROR Container restarted automatically

Output Screenshot

---

Configuration & Environment Variables

This service requires an OpenAI API key.

1. Create Environment File

Copy the example file:

.env.example → .env

2. Add Your API Key

Inside .env:

OPENAI_API_KEY=your_openai_api_key_here

3. Security Notes

• The .env file is excluded from version control.
• API keys must never be committed to the repository.
• In production environments, environment variables should be injected via:
  • Cloud secret manager
  • CI/CD pipeline secrets
  • Container environment configuration

Environment variables are loaded using python-dotenv.

---

Running Locally

Install dependencies:

pip install -r requirements.txt

Start the server:

uvicorn app.main:app --reload

Open:

http://127.0.0.1:8000

---

Failure Modes Considered

• Malformed JSON from LLM
• Missing structured fields
• Overconfident confidence scoring
• Truncated context hiding deeper root causes
• Increased inference latency

---

Token & Cost Considerations

• Logs are truncated to control token usage
• Deterministic preprocessing reduces prompt size
• Low temperature reduces retry frequency
• Future improvement: hash-based caching layer

---

Tech Stack

• FastAPI
• Async OpenAI SDK
• Pydantic
• Jinja2
• Python 3.10+

---

Positioning

This project demonstrates:

• Applied LLM system design
• Guardrailed AI integration
• Structured output enforcement
• Reliability-focused AI engineering
• Cost-aware prompt orchestration