NTU_cockroach

career agent flow
full architecture

Inspiration

Job searching is fragmented across spreadsheets, job boards, résumé files, and countless browser tabs. Most platforms rely on keyword matching and rarely explain why a job fits a candidate. We wanted to build a system that genuinely understands a person’s experience, remembers what matters, and helps them make better career decisions.

That idea became Northstar: an AI-powered career workspace built around persistent, transparent, and user-controlled career memory.

## What it does

Northstar turns a candidate’s résumé into structured career memory containing skills, experience, achievements, education, certifications, and preferences.

It uses this memory to:

Recommend live job opportunities based on the candidate’s actual experience.
Explain every match using evidence extracted from the résumé.
Learn from feedback such as “interested,” “wrong seniority,” or “hide this company.”
Cache repeated searches and refresh results when the candidate’s memory or preferences change.
Track job applications across different pipeline stages.
Display application and career insights through a unified dashboard.
Import and export application data using CSV files.
Capture job postings through a companion Chrome extension.

Users can inspect and delete individual memories at any time. Northstar never applies to jobs automatically—the candidate remains in control.

## How we built it

We built Northstar as a full-stack application using React and Vite for the frontend, with Node.js and Express powering the backend API.

CockroachDB Cloud acts as the primary database for users, sessions, applications, candidate profiles, career memories, search caches, and job feedback. Résumé facts are converted into 1,536-dimensional embeddings and stored using CockroachDB’s distributed vector indexing. A user-prefixed cosine vector index enables Northstar to retrieve the most relevant career evidence while maintaining strict per-user isolation.

We used the OpenAI Responses API to extract structured information from PDF résumés and discover current public job listings. The OpenAI Embeddings API converts career facts and search intent into vectors for semantic retrieval.

The project also includes secure authentication, HTTP-only sessions, CSV migration tools, structured logging, automated tests, a Chrome extension, and a containerized deployment setup compatible with AWS Lambda Web Adapter and Amazon ECS.

## Challenges we ran into

One of our biggest challenges was reliably extracting structured information from résumés. Résumés vary widely in formatting and often combine skills, responsibilities, and achievements in the same section. We addressed this with strict structured AI outputs, validation, confidence scores, and source metadata.

Another challenge was keeping vector retrieval secure. Similarity search had to return only memories belonging to the authenticated user. We solved this by using user_id as a vector-index prefix and applying user-scoped filters throughout the database layer.

Live job discovery created additional reliability problems. Search results can be duplicated, outdated, inaccessible, or linked to generic career pages. We implemented URL validation, normalization, deduplication, employment-type filtering, and an expansion pass when too few valid jobs were found.

We also had to carefully manage CockroachDB connections in serverless environments. We introduced bounded connection pools, configurable migrations, request timeouts, and smaller pool sizes for AWS deployments.

## Accomplishments that we're proud of

We are proud that Northstar became a complete, functional career platform rather than only an AI demonstration.

Our main accomplishments include:

Building persistent and inspectable career memory using CockroachDB vector indexing.
Producing explainable job recommendations grounded in résumé evidence.
Creating a feedback loop that updates preferences and invalidates stale results.
Enforcing strict data isolation across users, sessions, applications, memories, and feedback.
Combining résumé analysis, job discovery, application tracking, analytics, and browser capture in one platform.
Creating a read-only CockroachDB Cloud MCP demonstration for schema inspection, querying, and execution-plan analysis.
Providing database migrations, example datasets, deployment files, and complete setup documentation.
Delivering a passing automated test suite and successful production build.

## What we learned

We learned that effective AI memory must be transparent and editable. Users are more likely to trust an AI system when they can inspect what it remembers, see where the information came from, and remove anything incorrect or outdated.

We also learned that vector search is most effective when combined with structured metadata, identity filtering, result validation, caching, and explicit feedback. Embeddings provide semantic relevance, but the surrounding system determines whether recommendations are reliable and safe.

CockroachDB allowed us to store relational application data, authentication state, vector memory, feedback, and cached results in one distributed database. This simplified our architecture while maintaining consistency and scalability.

Most importantly, we learned that career AI should support people’s decisions rather than make decisions for them. User control and explainability became central parts of Northstar’s design.

## What's next for NTU_cockroach

Our next goal is to develop Northstar into a continuous career companion that supports users beyond a single job search.

Future improvements include:

Supporting multiple résumé versions for different roles.
Generating evidence-backed cover letters.
Providing personalized interview preparation.
Identifying skill gaps and recommending learning paths.
Sending alerts for newly discovered matching jobs.
Adding salary and career-progression insights.
Integrating more job sources and improving freshness validation.
Introducing email verification and password recovery.
Expanding recommendation-quality evaluations and observability.
Supporting multi-region deployment using CockroachDB and AWS.
Developing collaboration features for university career centers and mentors.

Our long-term vision is for Northstar to become a trusted career intelligence platform that grows alongside each candidate while keeping their data, memories, and decisions under their control.