Hello, that is Abdul Ashiq Abdul Malik, from Wichita State College. Coming from a background in computing, with expertise in information cleansing and extracting enterprise insights, I’ve all the time aspired to transition into the AI area. This undertaking has been an eye-opener, providing invaluable classes in agent-based techniques and generative AI.
On this weblog, I’ll share my perspective on the challenges I confronted whereas constructing my AI-powered journey assistant, designed particularly for vacationers planning journeys inside the United States.
Initially, my imaginative and prescient was to create a globally accessible journey assistant, however provided that this was my first undertaking within the area, I made a decision to start out with a extra targeted method. To streamline growth, I explored SmolAgents, a light-weight library that allows constructing highly effective AI brokers with minimal code.
1. Actual-Time Data Instruments
- Present time zone data for any US location
- Climate forecasts for locations
- Security insights for various cities
2. Journey Planning Capabilities
- Journey price range estimation based mostly on vacation spot, group measurement, length, and lodging preferences
- Lodging finder with price range filters (price range, mid-range, luxurious)
- Restaurant suggestions by delicacies, dietary desire, and worth vary
- Standard meals chain locator
- Attraction suggestions tailor-made to traveler pursuits
3. Transportation Planning
- Inter-city transportation choices
- Native transportation steerage
- Navigation between particular places
4. Technical Structure
- Powered by Qwen2.5-Coder-32B-Instruct
- Fallback to a Mock Mannequin when API limits are reached
- Gradio UI for seamless interplay
- Instrument-based structure utilizing perform calling capabilities
- DuckDuckGo integration for real-time internet searches
5. Mock Databases
- Intensive mock information for accommodations, eating places, and points of interest
- Completely different lodging tiers (price range, mid-range, luxurious)
- Detailed resort information: costs, rankings, options, and insurance policies
6. Person Expertise Enhancements
- Multi-step reasoning for complicated queries
- Conversational interface for pure interactions
- Help for picture era
After I first launched into this undertaking, I anticipated an easy method to programming and integration. Nonetheless, the journey turned out to be much more instructive than anticipated. My preliminary plan was to combine real-time information utilizing APIs, however I shortly realized that almost all complete journey APIs require paid subscriptions. Resort reserving APIs usually include vital utilization charges, whereas climate and transportation APIs comply with tiered pricing fashions.
Given these constraints, I thought of utilizing metadata as a substitute. This method had benefits — no dependency on exterior API uptime or sudden modifications, together with sooner growth cycles with out ready for API responses. Nonetheless, making certain the consistency of mock information proved to be a significant problem. It required cautious alignment, comparable to sustaining logical worth variations based mostly on resort high quality and making certain distances between places made sense.
For the AI mannequin, I selected Qwen2.5-Coder-32B-Instruct because of its sturdy function-calling capabilities, superior reasoning for complicated journey planning, and environment friendly code era. It additionally struck a stability between efficiency and value when in comparison with bigger fashions. Initially, I used Hugging Face’s API companies, however as a result of excessive inference prices, I shortly bumped into credit score limitations. This led me to implement a fallback mechanism, making certain that the system wouldn’t fail completely even when exterior companies grew to become unavailable.
I additionally encountered charge limits, significantly with RateLimitHandledDuckDuckGoSearchTool, which required retry logic and sleek degradation. To mitigate this, I experimented with smaller Qwen variants like Qwen2.5–7B-Instruct and Qwen2–7B, adopted by Mistral-7B-Instruct, Llama-3–70B-Instruct, and Claude-3-Sonnet. Whereas these supplied short-term aid, I wanted a extra sustainable answer. This led me to discover working fashions regionally utilizing Ollama, which was profitable however had limitations. I additionally thought of LM Studio, which allowed me to run giant inference fashions regionally, however shifting from Hugging Face integrations required substantial code modifications, making earlier API-based implementations out of date.
Architecturally, the instrument system underwent vital refinements. Initially designed for real-time API integration, it advanced right into a metadata-driven method because of exterior constraints. Creating a system the place 14 specialised instruments work together seamlessly required meticulous interface design, dependable perform calling, and cautious context administration between instruments.
For a greater person expertise, I targeted on pure dialog dealing with, making certain clean multi-turn dialogues and a standardized response format throughout all instruments. I selected Gradio UI for the interface, however encountered compatibility points, significantly parsing errors when dealing with perform calls. This grew to become a crucial space of focus for debugging and refinement.
The agent now operates with a hybrid method, counting on metadata however switching to internet searches when instruments fail to ship enough responses. Guaranteeing the system stays extensible — able to integrating new options with out main refactoring — posed further architectural challenges.
This undertaking highlighted the complexities of constructing a specialised AI assistant that balances technical functionality, person expertise, and the sensible limitations of contemporary AI fashions. Regardless of the challenges, the method has been a useful studying expertise, pushing me to refine my method at each stage.
Constructing this AI-powered journey assistant was each difficult and rewarding. It gave me real-world publicity to:
✅ API vs. metadata trade-offs
✅ Selecting AI fashions based mostly on value vs. efficiency
✅ Scalable tool-based AI architectures
✅ The significance of fallback mechanisms
This undertaking not solely strengthened my AI growth abilities but additionally gave me sensible publicity to real-world implementation hurdles.
I’m enthusiastic about what’s subsequent in my AI journey — keep tuned for extra! 🚀
💬 Have questions or suggestions? Drop them within the feedback!