Capstone focused on real-world conversational messiness.
Drive-Thru AI
End-to-end conversational AI system where the model acts as a drive-thru order taker, handling real-world, messy user input and maintaining context across a full ordering session.
1 / 3
Project summary
Conversational AI system simulating a fully automated drive-thru ordering experience.
End-to-end conversational AI system where the model acts as a drive-thru order taker, handling real-world, messy user input and maintaining context across a full ordering session.
What needed to be solved
Real-world conversations are noisy, non-linear, and full of interruptions and corrections.
- Most AI demos assume clean input.
- Most AI demos assume linear conversations.
- Most AI demos assume ideal user behavior, which doesn’t reflect real usage.
How it was built
Key implementation decisions, system behavior, and workflow structure.
- Designed a conversational system that maintains context across multiple turns.
- Handled interruptions and order changes.
- Filtered out irrelevant or extraneous conversation.
- Simulated real-world interaction patterns such as changing orders mid-sentence and adding or removing items dynamically.
- Focused on robustness, adaptability, and realistic conversational flow.
Tools and platform choices
Core technologies used in the project.
- Next.js
- Supabase
- TypeScript
- AI APIs (LLMs, prompt orchestration)
What mattered during implementation
Challenges, tradeoffs, and takeaways from the project.
Challenges / Tradeoffs
- Managing conversational state across multiple turns.
- Preventing the AI from drifting or losing context.
- Handling ambiguous or incomplete input.
- Designing prompts that are flexible but controlled.
Outcome / Lessons
- Real-world conversational systems require state management, not just prompts.
- AI must handle messy input, not idealized input.
- Robustness matters more than clever responses.
Related projects
More systems and applied AI work in the portfolio.
1 / 11
Coding Your Career
AI-powered learning platform with a custom LMS, AI teaching assistant, and automated content generation.
Custom LMS with AI teaching and grading workflows.
1 / 12
Sunday Go Lessons
Go learning platform combining structured lessons, problem training, and AI-powered game analysis using KataGo.
Structured teaching plus AI-powered game analysis.
1 / 6
Workflow IQ
Agent-based workflow system using DAG architecture to safely constrain and orchestrate AI agents.
2nd place at Tetrate buildathon.