TripMind: A Multi-Agent AI Travel Optimizer with $8 Cost and Comparative Study on LLM Fine-Tuning

This post will discuss the TripMind project, an autonomous multi-agent AI optimization system designed specifically for domestic travel in India. Its core goal is to help users save approximately 1000 rupees without compromising travel quality. The most prominent feature is its extremely low implementation cost (only $8 for data), and it conducts a comparative study using three different LLM fine-tuning strategies (SFT, Knowledge Distillation, Curriculum Learning). It also uses the MCP protocol to build a tool ecosystem and finally provides services via FastAPI. This thread will cover project background, technical architecture, fine-tuning strategies, evaluation results, cost analysis, and industry insights in separate floors.

• Original Author/Maintainer: aguru-venkata-saisantosh-patnaik
• Source Platform: GitHub
• Original Link: https://github.com/aguru-venkata-saisantosh-patnaik/tripmind_mcp_agents_llm
• Release Date: 2026-06-13
• Project Goal: Find the balance between price and quality for Indian domestic travel users, saving about 1000 rupees
• Coverage: 20 cities in India, 5 budget tiers
• Core Comparison: Performance comparison between three fine-tuning strategies based on the Llama 3.1 8B model and the non-fine-tuned baseline model

Multi-Agent System

• Architecture: Asynchronous pipeline architecture with a supervisor + three working agents (Analysis, Coordination, Optimization)
• Capabilities: Checkpoint recovery, concurrency control (max 3 concurrent trajectories), quality filtering (discard trajectories with loop calls/empty responses/fewer than 50 API calls)
• Agent Responsibilities:
  • Supervisor: Coordinate the entire process and manage MCP server connections
  • Analysis Agent: Collect data (routes, hotels, flights) and generate cost reports
  • Coordination Agent: Search for points of interest, restaurants, and propose alternatives
  • Optimization Agent: Integrate outputs, generate optimized itineraries and hub analysis

MCP Server 4 custom Model Context Protocol servers encapsulating external services:

Server	Port	Data Source	Tool Functions
routing_server.py	8001	OpenRouteService + Nominatim	get_route, geocode_city
hotels_server.py	8002	Overpass API (OSM) + haversine	search_hotels, search_flights
overview_server.py	8003	DuckDuckGo	search_pois, search_restaurants, web_search
pricing_server.py	8004	Internal pricing logic	get_cost_estimate, compare_prices

The project uses three QLoRA fine-tuning strategies for comparison:

1. SFT (Supervised Fine-Tuning)：Standard fine-tuning using 5000 synthetic itinerary pairs generated by GPT-4o-mini. Advantages: Low data cost and wide coverage, but may have synthetic data bias.
2. Knowledge Distillation：Distill knowledge from 500 multi-agent reasoning trajectories generated by DeepSeek V4 Flash. Advantages: Learn agent reasoning patterns, but limited data volume.
3. Curriculum Learning：Two-stage progressive training (simple samples first, then complex samples). Theoretically improves generalization ability, but requires careful design of the curriculum sequence.

Evaluation Framework

• Gold Test Set: 92 cases
• Evaluation Dimensions: 10 (structural integrity, semantic accuracy, LLM judge (DeepSeek V4 Flash), intent alignment, adversarial red team testing, etc.)

Key Results

• Structural Integrity: The output format standardization of fine-tuned models is significantly better than the baseline
• Semantic Accuracy: The knowledge distillation strategy performs best in understanding user intent
• Robustness: Passed 45 red team prompt tests to verify model stability

Data Cost

• GPT-4o-mini: $4 to generate 5000 synthetic training data
• DeepSeek V4 Flash: $4 to generate 500 multi-agent reasoning trajectories
• Total Data Cost: $8

FastAPI Service

• Endpoints: POST /optimize (submit optimization request), GET /results/summary (get result summary), model management endpoints (support dynamic switching of 4 models)
• Features: Async Ollama client, Pydantic model validation, auto-generated Swagger documentation

Project Value

1. Low-Cost Feasibility: Only $8 data cost proves that high-quality AI systems do not require huge investments
2. MCP Protocol Value: Standardized tool interfaces reduce integration complexity
3. Multi-Strategy Comparison: Provide data support for LLM training method selection
4. Modular Architecture: MCP server design improves maintainability and scalability

Industry Insights

• For startup teams/individual developers: Provide a complete reference implementation from data generation, model training to service deployment
• Technical Direction: The combination of multi-agent + LLM fine-tuning is an effective path in the travel optimization field