Ibraheem Wagih · Fraud & Financial Crime → Analytics
Projects / Agentic Systems

GoDo AI: Agentic Event Recommendations

Group of 3 Machine Learning for Business (MSIN0231) UCL MSc Business Analytics

A functional mobile MVP recommending evening plans to time-poor London professionals from calendar density, energy, budget and location. Built with Mahira Malikzada and Olena Brylinska. Design principle: a deterministic rules layer constrains what the LLM agent may suggest, and every recommendation is verified against a live source before display.

2 layers
deterministic rules + LLM reasoning agent
4 APIs
Ticketmaster, Skiddle, Eventbrite, Google Places
≤3 rounds
of agentic tool-calling per query
0
unverified venues shown to the user

The Working App

Two recommendation modes: a Daily Top-5 ranking engine, and Scout, an agentic chat agent doing up to three rounds of tool-calling against live event APIs, streaming results to the React Native client over SSE. Suggestions carry transparent reasoning, and booking writes back to the user's calendar.

GoDo AI Suggested feed on a phone
The MVP's Suggested feed: daily curated picks with Interested / Not Interested feedback and calendar-synced booking.
GoDo AI multi-phase AI selection pipeline
The two recommendation modes: deterministic-first Daily Top 5, and the agentic Scout chat with tool-calling, SSE streaming and persistence.

Architecture and Guardrails

Layer 1 is rules: Burnout Guard scoring maps calendar density to recommended activity intensity, with hard budget and time filters. Layer 2 is a Claude-powered reasoning agent that interprets context and writes explanations, operating only inside what the rules allow. Every venue is verified against a live API before display, so the agent cannot hallucinate a restaurant that does not exist. Privacy was designed in from the start: read-only Free/Busy calendar access and mutual-interest gating before any social feature activates.

GoDo AI two-layer architecture slide
Two layers, one rule: deterministic guards decide what is in scope, the LLM handles nuance inside those bounds.

My Contribution

Within the team of three, I worked on the two-layer architecture design, the two recommendation modes including the tool-calling agent loop, the API-verification grounding approach, and the privacy guardrail definitions. I also contributed to the pitch deck and the MVP walkthrough video.