Skip to content

Capstone · Submitted

VisaPros

Google × Kaggle 5-Day AI Agents Intensive capstone. Agentic visa eligibility advisor using Google ADK to run six parallel country agents (UK · CA · US · AU · NO · DE) in a collaborative agent-to-agent (A2A) network with Google Search grounding.

Role
Creator · Solo build · Public repo
Period
Nov — Dec 2025
Stack
Python · Google ADK · A2A Hub-and-Spoke · A2A servers (uvicorn) · and 3 more
Place
public on GitHub
Submitted to the Google × Kaggle Agents Intensive capstone (2025) — "An ecosystem of AI solution to Visa Eligibility for Embassies and Individuals."

Submitted to the Google × Kaggle Agents Intensive capstone (2025) — "An ecosystem of AI solution to Visa Eligibility for Embassies and Individuals."

Google × KaggleCapstone (Submitted)

Anyone considering migration faces the same wall: dozens of countries, dozens of visa categories, dozens of evolving rule sets, and zero way to compare your actual eligibility side-by-side. Generic chatbots hallucinate immigration law; consultants charge for what should be a 20-minute parallel evaluation.

01 / 04

The problem

VisaPros treats each destination as a sovereign expert.

You upload your documents once; six country specialists read them in parallel and report back what you qualify for, what's missing, and which route is the highest-yield next step.

02 / 04

What I built

  • Google ADK Agent-to-Agent (A2A) hub-and-spoke architecture. A central orchestrator using Google ADK parses documents sequentially, then fans the structured profile out to six country agents that evaluate independently, each served as its own A2A server.
  • Six destination specialists. UK · Canada · US · Australia · Norway · Germany — each agent knows its own visa categories (tourism, business, study, healthcare, work) and applies them against the parsed profile.
  • Real-time grounding via Google Search. Country agents validate against current immigration rules at evaluation time, not stale training data — critical because immigration policy drifts monthly.
  • Strategic advisor synthesis layer. Ranks countries by eligibility yield, flags the "low-hanging fruit" path, and recommends the smallest set of additional documents that unlock the most options.
  • Structured JSON document extraction. Passport, education certificates, work history, financials — all normalised into a profile schema the country agents consume identically.
03 / 04

What I learned

The first architecture was a single linear agent that walked all six countries serially.

It was correct, slow, and brittle — one country's edge case stalled the whole pipeline.

Splitting into parallel country agents, each its own A2A server, was the unlock. Accuracy lives in the sequential parser; speed and resilience live in the parallel agents; judgement lives in the strategic advisor. Three separate concerns, three separate components — exactly the boundaries a multi-agent system is supposed to have.

04 / 04

Capstone context

Built and submitted as the capstone project for the Google × Kaggle 5-Day AI Agents Intensive (2025) — a programme focused on production-grade agentic systems: multi-agent orchestration, grounding, evaluation, and tool use.

VisaPros is the artefact: a complete A2A agent mesh, real-time grounding, and a synthesis advisor — shipped solo, end to end, and public on GitHub.

Decisions and why

  1. Sequential parsing, parallel evaluation, synthesised advice

    WhyDocument extraction is accuracy-critical and must stay deterministic; country evaluation is independent and parallelisable; final recommendation needs cross-country judgement. Three distinct workloads, three distinct components.

  2. A2A microservices over a monolithic agent

    WhyA single linear agent stalled on one country's edge case. Splitting into six independent A2A servers means each country can fail, retry, or update its rules without blocking the others.

  3. Google Search grounding at evaluation time

    WhyImmigration policy drifts monthly. Baking rule data into the model weights guarantees stale answers. Live grounding keeps every evaluation current — and forces the agent to cite its source.

Technical highlights

  • 01

    Agent-to-Agent (A2A) hub-and-spoke orchestration across six country specialists.

  • 02

    Gemini 2.5 Flash Lite for low-latency parallel inference across all destinations.

  • 03

    Each country agent is its own A2A server, independently deployable.

  • 04

    Google Search grounding for real-time immigration rule validation.

  • 05

    Sequential document parsing → parallel evaluation → strategic synthesis (three-stage hybrid pipeline).

Impact

  • Replaces multi-hour, multi-consultant migration research with a single parallel evaluation across six destinations.
  • Surfaces the "highest-yield document to acquire next" — turns vague aspiration into a concrete next step.
  • Designed for migrants and remote workers from the Global South — the audience most under-served by existing visa-tech tooling.
Outcomes

By the numbers.

6
Country agents in parallel
Any
Visa type — tourism, study, work, healthcare
Visit project

See it in the wild.

This case study covers how VisaPros was built — these links open what ships today: the live product, repo, demos, and supporting material.

↗ Next case study

Vimalgo

Vimalgo is the AI-driven headhunting and recruitment management system ELO built for Cannella & Partner Solutions AG, a Swiss staffing firm in Möhlin. It automates the part of recruitment that used to be pure manual reading — parsing CVs into structured candidate profiles, keeping a job database current from external portals, and surfacing high-potential candidate–job matches — then tracks each application end to end from first contact to placement. Fauzul was the main driver of the AI demo that opened the engagement: showing a working AI matching flow early enough that the client chose to bet on the AI promise rather than wait for it, a bet the platform has since delivered on.