Career software · Case study
A closed-loop AI workspace for the job search.
A product connected direct-source role discovery, evidence-grounded fit ranking, truthful resume tailoring, application assistance, and communication tracking around one durable job record. Different AI lanes could be evaluated and routed independently instead of sharing one model decision.
- Work
- Product, AI, data, and platform engineering
- System
- TypeScript, Python, Postgres, AWS
- Focus
- Closed-loop workflows and model evaluation
The challenge
A job search is a workflow, not a chatbot session.
Finding a role, deciding whether it fits, presenting experience accurately, applying, and tracking what happens next are separate tasks with different evidence and risk. If each AI feature keeps its own state, the user has to reconcile the product’s memory by hand.
The product needed one durable loop: new evidence should improve the same job record, and each model-assisted step should remain inspectable before it influenced the next.
What the work involved
Several AI lanes. One product truth.
The architecture kept the user’s source material and each job’s evidence at the center.
Found roles from first-party sources.
Source adapters, normalization, canonicalization, and deduplication brought jobs into a common model while preserving where the evidence came from and how fresh it was.
Combined fast signals with deeper judgment.
Behavior, semantic similarity, experience fit, and explicit preferences handled ordinary comparisons. More expensive model reasoning could be reserved for ambiguous cases and grounded in the actual posting.
Made tailoring truthful by construction.
The resume workflow worked from structured source experience, retained revision history, and treated unsupported claims as failures—not improvements to the prose.
Closed the loop after the application.
Email and calendar signals attached to the same job record, where stage, interview details, and next actions could be reviewed without creating a second competing source of truth.
Evaluation system
Quality meant something different in every lane.
Could the system distinguish an official live role from an aggregator or stale copy?
Did the score honor dealbreakers and explain itself from the role and candidate evidence?
Did the artifact sharpen real experience while retaining every factual boundary?
Did the system connect a messy thread to the right job and propose the right next step?
What transferred
Closed-loop AI needs a canonical record, not a longer conversation.
A durable domain object gives every model, tool, and person a shared place to read and write evidence. It also makes evaluation more honest: the system can be judged on the state it creates and the decisions it supports, not only on the text of one response.