Career Atlas

QR For Live Demo

💡 Inspiration

We are university students with a passion for AI, but even that passion came with a paralysing question: should we become a Machine Learning Scientist, an ML Engineer, an MLOps Engineer, or an AI Engineer? These titles overlap, diverge, and evolve constantly. No one around us had a clear answer, and every Google search sent us down a different rabbit hole.

We looked around and realised we weren't alone. Most of our peers were in the same fog, not laziness, not lack of ambition, but a genuine absence of direction. We fumble. We pick courses that sound right. We apply for internships we're underqualified for and wonder why. The career journey for most university students is less of a plan and more of a series of educated guesses.

The tools that exist today don't help enough. Job boards show you what companies want, not what you have. Career fairs give generic advice that doesn't know your background. Asking friends or seniors is valuable, but they're confined to their own experiences and biases.

What we needed wasn't more information, we needed a personalised mentor: one that could read what we've actually built, understand where we want to go, and map the gap between the two with the breadth of knowledge no single person could have.

✨ What it does

Career Atlas is a personalised AI career companion for university students and fresh graduates who don't know where to go or how to get there.

It works in three stages:

Stage 1: 🗂️ Build Your Living Portfolio

Upload your resume, certificates, and transcripts. A multi-agent AI pipeline reads your actual documents — not just your self-reported claims — and constructs a verified skill portfolio, cross-referencing your career experience, project work, and academic background. Skills are tagged by how they were evidenced: officially (certified) or AI-inferred, so you always know how strongly each skill is backed.

Stage 2: 🎯 Identify Your Goal

Not sure where you want to go? Career Atlas AI interviews you: conversationally, like a mentor. It adapts to three types of users: those with no idea, those with a general direction, and those with a specific dream job or job description in mind. It asks the right questions, narrows down your preferences, and helps you commit to a concrete, realistic career target, grounded in your actual skills, not wishful thinking.

Stage 3: 🗺️ Get Your Personalised Roadmap

Once your goal is set, a three-agent pipeline goes to work:

Agent 1 fetches live market skill requirements for your target role via Google Search grounding, reflecting what employers actually want today
Agent 2 performs a deterministic gap analysis, comparing what the market requires against what your portfolio proves
Agent 3 builds a 3-phase, personalised upskilling roadmap, prioritised by what matters most, with portfolio-building actions at every step

The result isn't generic advice. It's your gap, your roadmap, grounded in real evidence and live market data.

🔧 How we built it

Career Atlas was built by a team of four, each owning a distinct layer of the stack:

Casimir Chiong — Architect & All-Rounder, responsible for system design and cross-cutting concerns
Tharmu — Frontend (Next.js, TypeScript)
Jee See — Backend (Supabase, API routes)
SL Kok — AI Engineer (Vertex AI pipelines)

We followed a deliberate build order that kept us from over-engineering before validating:

Architecture first: defined the data flow, agent boundaries, and API contracts before writing a single component
Hard-coded prototype: validated the core idea with static inputs to test whether the AI outputs were meaningful
Living Portfolio: replaced static data with a real multi-agent extraction pipeline that reads actual user documents (resumes, certificates, transcripts) and builds a verified skill portfolio
Roadmap: the final and most complex layer: a three-agent pipeline that fetches live market skill requirements via Google Search grounding, performs a deterministic gap analysis, and generates a personalised 3-phase upskilling plan

For AI infrastructure, we chose Google Vertex AI with Gemini 2.5 Flash (conversational goal identification) and Gemini 3.5 Flash (structured roadmap generation). This was made possible by the Google Developer Program premium credit through the Google AI Pro Plan Student Free Trial, which also made the decision an obvious one. We're grateful for the access.

😤 Challenges We Ran Into

Designing for Human-in-the-Loop Without Breaking the Flow

The hardest architectural challenge wasn't making the AI work, it was deciding where humans should intervene. Career Atlas isn't meant to be a black box that hands you a roadmap and disappears. We wanted meaningful human checkpoints: confirming your goal before generating a roadmap, reviewing your skill portfolio before committing it, adding context the AI might have missed.

Defining these touchpoints cleanly, without making the product feel clunky or paternalistic, required multiple iterations. Too much friction and users skip the step. Too little and the AI runs ahead with bad assumptions.

Getting the AI to Stop Making Things Up

Hallucination was our most persistent technical adversary. An AI that confidently attributes a skill you don't have is worse than no AI at all, it poisons every downstream output.

Our response was layered: we enforced structured output schemas at every agent boundary, built deterministic TypeScript guardrails outside the LLM (so the model cannot override them), required every extracted skill to cite the exact source document filename, and added a cross-referencing directive that forces agents to validate claims against provided evidence before extracting anything. If a skill has no supporting evidence, it gets dropped, deterministically, not by asking the model nicely.

Narrowing the Scope

The hardest non-technical challenge was knowing what not to build. Career guidance is an enormous problem space. We could have built interview prep, salary benchmarking, networking tools, application tracking, etc. The discipline of continuously asking "does this serve the core user need?" was uncomfortable but necessary. Every feature that didn't directly serve the path from confusion to concrete plan got cut.

The good news: our bugs were small. Nothing in beta testing took more than 15 minutes to fix, thanks to early architecture design.

🏆 Accomplishments that we're proud of

We Built Something We Actually Use

The most honest measure of whether a product works is whether its creators use it. We do. When we ran our own resumes through Career Atlas, we got clarity we didn't have before. We have a clearer sense of where we stand, which skills we've genuinely evidenced, and what we'd need to close the gap to roles we'd actually want.

What would have taken days of research, trial and error, and second-guessing now takes under an hour, with the same level of clarity and confidence at the end of it. That's the accomplishment we're most proud of.

The Clarity Moment

During testing, we each went through the goal identification interview ourselves. For a team that includes people genuinely unsure about whether to pursue ML Engineering, MLOps, or AI Engineering, that conversation with the AI was unexpectedly useful. It asked the right questions, pushed back on vague answers, and narrowed us down to a specific profile. Watching the roadmap generate from our actual portfolios and confirm that clarity step-by-step was the moment we knew we'd built something real.

A Simple Architecture That Actually Works

There's a temptation in hackathons to over-engineer. We resisted it. Three clean pipelines. Clear agent boundaries. Deterministic guardrails that live outside the LLM so the AI cannot hallucinate its way past them. A multi-agent extraction system that verifies skills across three independent evidence streams before a single one makes it into your portfolio.

Simple but not naive. The architecture held up under real use, real documents, and real edge cases. No catastrophic bugs. No last-minute rewrites. That's something we're genuinely proud of as a team.

🎓 What We Learned

Reliable AI Is an Engineering Problem, Not a Prompt Problem Getting a generative model to do what you want, once, is easy. Getting it to do what you want reliably, across diverse inputs, edge cases, and adversarial prompts, is genuinely hard. We learned that reliability doesn't come from a better prompt alone. It comes from the system around the model: structured output schemas, deterministic validation layers, evidence requirements, and guardrails that the LLM simply cannot override. The model is powerful but it needs a well-designed cage.

The Human Must Always Be in the Loop The deeper we went into building Career Atlas, the more we became convinced of one thing: AI is a decision support tool, not a decision maker. It can surface options you wouldn't have considered, compress days of research into minutes, and ground your choices in evidence. But the commitment to a career goal, to a roadmap, to the next step, must always be yours. We designed every checkpoint in the product around this principle, and it made the product more trustworthy, not less powerful.

The Question Is No Longer "Can We Build It?". It's "Should We?" This may be the most important lesson of all. With modern agentic coding tools, the barrier to building has collapsed. A small team of students can ship a multi-agent AI pipeline in a week! When almost anything is buildable, the hard discipline shifts entirely to product thinking: Is this the right problem? Is this the right solution? Does this actually help the user, or does it just technically work? Those questions are harder than any engineering problem we faced and they're the ones that will matter most going forward.

🗺️ What's Next for Career Atlas

Immediate: Make What We Have Trustworthy

We're not chasing features. The next priority is hardening what we've built — more testing across diverse user inputs, more guardrails, and more honest evaluation of where the AI still falls short. A roadmap that works 80% of the time isn't good enough when someone is making real decisions about their career. We'd rather ship something reliable than something impressive.

Beyond Students: Only With the Right Voices

Career Atlas was designed by students, for students. That was deliberate — we built from lived experience, and that specificity made the product sharper. But we recognise the limits of that. Expanding to career switchers, mid-career professionals, or returners to the workforce isn't just a product decision — it's a research decision. We'd need those users to help us build it. The worst outcome would be building a "general" career tool that's actually just a student tool with the labels changed.

Growth, if it comes, will be community-led — not feature-led.

The Bigger Vision

At its core, Career Atlas is about one thing: helping people understand themselves and navigate what comes next — with confidence, not guesswork.

If that works for a student choosing between ML Engineering and MLOps, why couldn't it work for a 40-year-old considering a pivot? Or a retiree exploring what to do with the next chapter? Or someone just starting to figure out who they are?

The vision, in its most ambitious form, is simple: personalise the human journey — from the first career question to the last. Not with AI making decisions for you, but with AI helping you make better ones, at every stage.

We start with students. But the map has no boundaries.