Every AI career platform seems to follow the same playbook.
Upload your resume.
Get an ATS score.
Match it against a job description.
Generate a cover letter.
Practice interview questions.
Chat with an AI assistant.
We built many of those features too.
Then we stopped and asked ourselves:
If everyone can build these features with an LLM in a weekend, what are we actually building?
That question completely changed the direction of our product.
The Problem We Kept Seeing
Candidates don't usually struggle because they don't have another resume builder.
They struggle because nobody can answer a much bigger question:
"What should I do next to become more employable?"
Should they:
learn Docker?
build an AI Agent?
improve system design?
deploy a project?
learn RAG?
focus on MCP?
switch to backend engineering?
Most tools generate documents.
Very few help candidates make better career decisions.
We Started Removing Features
Some of our AI features looked impressive in demos but weren't creating long-term value.
We experimented with:
AI Career Chat
AI Practice Modules
Generic AI generators
People tried them once.
Then never came back.
That taught us something important.
Users don't necessarily want more AI.
They want better guidance.
Instead of More Features, We Built Two AI Agents
Rather than adding another generator, we introduced two specialized agents with different responsibilities.
🧠Career Strategist
The Career Strategist continuously analyzes:
resume
target role
hiring market trends
skill gaps
portfolio
market fit
Its responsibility isn't generating content.
Its responsibility is answering:
"How employable am I today, and what should I improve next?"
🏗️ Portfolio Architect
The Portfolio Architect receives the Strategist's findings.
Instead of suggesting random GitHub projects, it designs a portfolio roadmap that aims to maximize employability.
It considers:
hiring demand
missing skills
project difficulty
expected career impact
logical learning progression
Instead of saying:
Build Project A.
It explains:
Build Project A first because your profile already demonstrates backend experience, but lacks deployment and production architecture. Completing this project is expected to improve your market fit for your target role.
The Interesting Part
The agents don't work independently.
They collaborate.
Resume
↓
Career Strategist
↓
Market Analysis
↓
Skill Gap Detection
↓
Portfolio Architect
↓
Personalized Roadmap
↓
Interview Coach
↓
Resume evolves
↓
Career Strategist runs again
Rather than isolated AI tools, we're experimenting with an AI system where each component has a single responsibility and shares context with the others.
Some Challenges We Ran Into
This turned out to be much harder than prompting an LLM.
A few problems we had to solve:
Preventing hallucinated career advice.
Keeping recommendations consistent across multiple agents.
Maintaining context from the candidate's active resume.
Explaining why a recommendation was made.
Making the UI feel like an autonomous system rather than a chatbot.
Balancing deterministic logic with LLM-generated insights.
We're still iterating on all of these.
What We Learned
The biggest lesson wasn't technical.
It was product-related.
People don't really need another AI that writes text.
They need an AI that helps them make better decisions.
That's the direction we're exploring.
Instead of asking:
"Can AI write this?"
We're asking:
"Can AI continuously improve someone's career strategy?"
Where We're Going Next
These agents are still in their early stages.
Next, we're focusing on:
richer long-term memory
autonomous monitoring of hiring trends
proactive recommendations instead of one-off reports
tighter collaboration between agents
measurable career progress over time
The goal isn't to replace human decision-making.
The goal is to help candidates understand a rapidly changing hiring market and adapt faster.
I'd genuinely love feedback from other developers and builders.
Does this architecture make sense?
Have you experimented with collaborating AI agents?
What challenges did you run into?
I'm especially interested in hearing how others are approaching persistent context, agent collaboration, and long-term memory in AI products.
Top comments (6)
I like the shift from content generation to decision support. One thing I’d be careful about is optimizing for multiple agents instead of measurable outcomes. Whether it’s one agent or five matters less than whether the recommendations actually improve interview rates, offers, or time to hire. Those feedback loops will probably become the real differentiator.
That's a great point. The number of agents isn't the end goal—it's just an implementation detail. The real objective is exactly what you mentioned: improving measurable career outcomes rather than simply generating better recommendations.
One of the challenges we're thinking about now is how to close that feedback loop so the system can learn from outcomes over time instead of optimizing only for intermediate signals. Thanks for bringing that up.
I think the hardest part isn’t closing the feedback loop—it’s deciding what counts as success. An interview invitation, an offer, salary growth, retention after six months, or even candidate satisfaction can all point in different directions. Optimizing for the wrong metric can easily lead the agents to game the system instead of genuinely improving careers. How are you thinking about defining success?
That's exactly the tension we're working through. For now, we're keeping it simple — just asking users "was this recommendation helpful?" with an optional outcome dropdown (got interview / got offer / skill improved).
The thinking is: collect data first, weight it later. No composite score yet — just raw signals that we can feed back into the agents once there's enough to be statistically meaningful.
Curious how you'd approach the weighting problem once there's enough data.
I like the idea of collecting raw signals first, but I’d avoid treating all outcomes as equal.
“Got an interview”, “got an offer”, “salary increase”, and “felt helpful” measure very different things. Some are objective outcomes, others are subjective perceptions, and they often conflict.
I’d model them as separate dimensions rather than collapsing them into a single score. Let the agents optimize for the user’s explicit objective (e.g. maximize interview rate, maximize compensation, minimize time to hire, or maximize long-term retention) instead of optimizing for one universal definition of success.
Otherwise, you risk building agents that optimize for the easiest metric rather than the user’s actual career goals.
That's a great framework — modeling outcomes as separate dimensions with user-selectable objectives is exactly where this should go.
For now though, we're in data collection mode. The dropdown exists, the table is live, and once we have enough signals we can build the multi-objective optimization layer you're describing. First, prove users report outcomes. Then, let them choose which outcome matters most.
Appreciate the push toward rigor — this is the kind of thinking that separates a toy from a real product.