Syncareer / Engineering case study
Career tools, with context.
A career platform connecting saved opportunities, CV tailoring and interview preparation in one application record.
- My contribution
- Product & application engineering
- Project state
- Live product
- Implementation
- Grounding revision in source
- Deployment
- Revision not yet redeployed
Stack TypeScript / React + Vite / Supabase Postgres + RLS / Edge Functions (Deno) / Lovable AI gateway
02 / Project visual
The product, as it runs.
Three real captures from the live platform. The grounding revision isn’t in them yet.
Readiness, applications, CV strength and interview progress, drawn from the same application records.
Saved opportunities feed CV tailoring and interview practice. Match scores are product UI, not a hiring probability.
Profile context and quick actions in the existing platform.
03 / Problem
The assistant saw a bullet. Not the evidence.
Before the change, the assistant received the selected bullet and nothing else. It could polish the wording. It couldn’t know the requirement the bullet was meant to answer, or tell a job keyword from candidate evidence. Replies came back as text plus broad source ids, and new claims were never checked.
As documented in the project’s own guidance notes.
04 / Contribution
The application layer, not the model.
One path inside the product: the CV assistant that rewrites a selected bullet for a selected job requirement. The problems lived around the model — context, validation, consistency — so I worked on the application layer that surrounds it.
- The contract — what a request may carry, and what a reply must cite before the interface can use it.
- The checks — validation and grounding between the model and the interface, not inside a prompt.
- The review surface — accept, edit, reject or undo stays with the person. Nothing reaches the saved CV on its own.
05 / Key decisions
Fix the contract, not the model.
This comparison is the whole investigation. The Lab runs one reply through the gate.
A request/response contract: allowlisted context in, validated and cited proposals out, and nothing applied automatically.
Four engineering decisions
Bounded, task-family server prompts
Ten tasks, each with its own server-side prompt built only from the supplied context. Job text, CV text and the instruction are treated as untrusted data.
Allowlisted context with size limits
Context enters only as allowlisted items with fixed provenance: 12 items, 8,000 characters each, 24,000 in total, 2,000 for the instruction.
Citation enforcement
Missing grounding is a refusal, not a warning. The request is checked before the model is called, the reply against the ids that request supplied.
Factual-risk checks and explicit review
New numbers, unsupported skills, employers presented as experience, coursework presented as employment, participation presented as a win, and generic filler all get flagged. A person decides what happens next.
The repair stayed in the existing career-guidance seam: React + Vite on Supabase, hosted on Lovable Cloud, no custom API server. Authorization is enforced server-side; the router’s guards are UX only.
The gate is on the Lab page. Write a bullet, choose what it cites, and watch the check decide. One example passes with four invented claims in it, which is the point.
06 / Result & status
A repaired contract. A release still pending.
A bounded request and cited-proposal contract is in source. The live model’s grounding has not been measured.
- Live product
- Deployed. The existing platform is public at syncareer.me, and the captures above come from it.
- Grounding revision
- In source, awaiting deployment. The live platform does not run it yet.
In source
- Request contract, task/provenance allowlists and size limits contract.ts
- Runtime validation of model JSON, with named failure codes parseModelProposal
- Requirement/evidence citation enforcement for cv.rewrite_bullet handler.ts
- Fail-closed ordering, with the reservation released on every failure handler.ts
07 / Limitations
What this does not show.
Not verified
The repository shows a repaired contract. It doesn’t show that the model got better.
- No live-model evaluation and no captured before-and-after model output
- No measured failure rate, latency, or output-quality score
- No authenticated canary — the four workflow families have not been run against the live model
- Stricter server prompt and citation enforcement are not live yet: the changed edge function still needs a Lovable Cloud deployment
- Interview generation and CV parsing run in deployed-only functions, whose prompts and quotas are not auditable from the repository
The project’s own notes score five fixed examples against a six-part rubric: 8/30 for the old contract, 29/30 for the revised one. Those are marks on fixtures that compare two contract versions, not live-model scores.
08 / Sources
Check the record.
- Live productsyncareer.meThe deployed platform the captures come from.
- Source repositoryGitHubContract, handler, prompts and the Deno tests.
- Contract handoffdocsThe project’s own record of what changed, what was tested, and what was not.
- Reliability LabThis siteStage-by-stage source inspection, plus the labelled fixtures.
Files cited: contract.ts, handler.ts,parseModelProposal, index.test.ts.