SofIA is a bilingual AI companion that helps Latino newcomers to Canada navigate their professional job transition, from understanding why reaching out to
other professionals matters, to actually enjoying it. Operating natively through WhatsApp, SofIA meets users where they already are, with no app to
download. Built through the Aggregate Intellect Agentic AI BuildCamp, informed by 9+ user research interviews with newcomers across Canada.
Objective
Latino newcomers often arrive in Canada with strong professional backgrounds but face a specific emotional barrier: fear and overwhelm around professional
conversations in an unfamiliar culture. The goal was not to build another job board or resource directory, but to address the emotional and practical gap
between knowing networking matters and actually doing it, in the user's own language, at their own pace.
What SofIA does
Guides users through preparing for and reflecting on professional conversations ("coffee chats")
Tracks contacts and follow-ups so nothing falls through the cracks
Helps users set career goals with encouraging progress tracking from onboarding to job offer
Surfaces regional newcomer resources tailored to the user's city
Responds naturally in English or Spanish based on user preference
Provides a web dashboard for reviewing contacts, progress, and conversation history
What SofIA explicitly does not do: replace mentors or career coaches, answer immigration questions, or substitute for mental health support.
Solution Architecture:
Component
Type
Role
Webhook Router
API
Receives incoming WhatsApp messages via Twilio, validates requests, routes to conversation handler
Claude Agent
Agent
Conversation orchestration, response generation, bilingual tone management, boundary enforcement
Session Manager
Service
User state persistence, onboarding flow control, conversation continuity across sessions
Contact Tracker
Service
Logs professional connections, records conversation outcomes, surfaces follow-up reminders
Goal Engine
Service
Career milestone tracking from onboarding through job offer, progress summarization
Web Dashboard
Frontend
Jinja2-rendered user profile, contact history, and learning records — accessible via browser
Langfuse Observer
Utility
LLM tracing, evaluation scoring, and quality monitoring across all production conversations
Firestore DB
Database
Persistent storage for user profiles, conversation history, contacts, and goal state
User Research:
9 in-depth interviews with Latino newcomers in Canada: 5 still navigating their job transition (Track A), 4 who had successfully transitioned to roles
in their field (Track B). Participants came from Colombia, Peru, Mexico, and the Dominican Republic, with backgrounds in healthcare, academia,
engineering, law, and technology.
Key findings:
All 9 participants arrived with the wrong mental model of professional outreach, the LATAM model (networking = who can get you the job?) vs. the
Canadian model (networking = visibility over time). Not one had been explicitly taught the difference. One participant executed 20 conversations in 30 days
after receiving a single clear explanation.
7 distinct forms of the "I don't want to bother them" barrier emerged across participants, from cultural inhibition and cold outreach fear, to social
anxiety developed in Canada, to a specific block around asking for help from warm contacts after the connection is already made.
3 of 5 still-searching participants attended professional events, collected LinkedIn connections, and never sent a follow-up message, believing
attendance was the work. The relationship starts after the event, not at it.
Emotional sustainability was the unanimous #1 gap for participants inside an active search: "La parte de la carga emocional es la más importante — sin
energía para seguir, el resto pierde importancia." — P3
All 4 participants who had successfully transitioned credited volunteering, informal mentorship, or sustained community involvement (not job boards), as
the primary path to landing their role.
Tools
Python
Claude (Anthropic)
FastAPI
Twilio (WhatsApp)
GCP Cloud Run
Firestore
Langfuse
Achievements
Deployed production-grade WhatsApp agent on GCP Cloud Run with zero app install required for users
Built bilingual conversational experience that switches naturally between English and Spanish
Designed and ran LLM-as-judge evaluation pipeline across 14 test cases, identifying 2 systematic response gaps
Conducted 9+ user interviews to validate core assumptions before writing a line of code
Implemented Langfuse observability for real-time LLM tracing and quality monitoring in production
Validated product with 2 real users actively using the tool