Ashir Rashid

Ashir Rashid

ML Engineering & Applied AI Research

AiXplore

The project

Hackathon project. Multi-agent travel assistant for the MENA region, built on the aiXplain platform. You describe a trip in plain language (Arabic included): budget, time limit, current location, preferences. Three complete itinerary options come back with restaurants, attractions, commute times, and taxi fare estimates for each stop.

5
Agents in sequence
3
Complete itinerary options returned
10+
Restaurant and attraction options each
AR / EN
Arabic input and output supported

Architecture

Five agents in sequence:

1
Input parser
Pulls budget, time limit, current location, final destination, meal type, and food and attraction preferences out of natural language.
Arabic input → Microsoft Translation first
▾
2
Restaurant search
Finds restaurants near the current location via Tavily and Google Maps.
10+ options: coords, distance, commute, price/person, rating, dish pick
▾
3
Attraction search
Same structure as restaurant search, tuned for places to visit.
10+ options: coords, entry price, distance, taxi fare estimate
▾
4
Itinerary generator
Takes the constraints and both search outputs and builds plans that stay within time and budget, account for commute and taxi cost, and end at the final destination.
Builds 3 plans
▾
5
Itinerary inspector
Reviews each plan for accuracy, re-checks taxi fares and commute times against Maps data, and translates the output.
Output translated to Arabic

Key design decisions

Sequential over parallel

Parallel search agents were tested. The handoffs got messy at coordination time, so the sequential pipeline stuck: the generator got cleaner inputs and more predictable output.

Inspector as its own agent

One agent that generates then validates its own itinerary tends to rationalize its errors. A separate inspector reviews the plans cold and fixes taxi fares and timing without that bias.

Google Maps for ground truth

Tavily gives approximate distances. A custom utility model wraps the Maps API as a callable tool, so agents get real driving times, distances, and place details instead of guessing.

Tavily ≈ approxMaps = ground truth

Technologies: Python aiXplain Google Maps API Tavily Microsoft Translation

Concepts / Algorithms: Multi-Agent Orchestration LLM Tool Use Agentic Pipelines Arabic NLP