ApplyVerdict
An honest fit score for remote jobs. Every claim traced to your CV.
Paste your CV and a remote job post. In a few seconds you get:
- Can you even apply from your country? Many "remote" jobs only hire in some regions. The verdict quotes the line in the post that decides it.
- Which requirements you really meet, each with the sentence from your CV that proves it. Claims the model can't back with a real quote are dropped.
- A fit score computed in code, not by the model's mood: the model only judges each requirement, and a fixed rubric turns those judgments into a number and a band.
- A cover letter, a cold DM and stronger CV bullets that only use facts from your CV. An honesty check highlights any name or number it can't find in your CV.
No signup. 3 free analyses a day.
Live: https://jobpilot-web.buildli731.workers.dev · Evals: /evals
| Landing | Evidence trace (live run) | Drafts with honesty check |
|---|
| | |
Why I built it
I'm a junior developer in Morocco looking for a remote AI-engineering job. Half the "remote" posts I read turned out to be US-only or EU-only, and every AI job tool I tried either inflated my fit or wrote cover letters with experience I don't have. ApplyVerdict is the tool I wanted: it says no when the answer is no, and it shows its evidence.
How it works
flowchart LR
A[CV + job post + country] --> B[Region rules<br/>country → regions]
B --> C[LLM: structured judgments<br/>verdict, requirements,<br/>verbatim quotes]
C -->|streamed partials| UI[Evidence trace UI]
C --> D[Evidence check in code<br/>every quote must exist<br/>in the CV or the post]
D --> E[Scoring rubric in code<br/>must / nice / seniority → 0–100 + band]
E --> F[Signed analysis token<br/>HMAC, 2 h]
F --> G[LLM: drafts]
G --> H[Honesty check<br/>names & numbers not in the CV are flagged]
H --> UI
- The model judges, code decides. The LLM returns Zod-validated judgments with verbatim quotes. Code verifies each quote against the inputs (unverifiable claims are downgraded), then a deterministic rubric computes the score. The UI never shows a score the server didn't compute.
- Streaming end to end. Both endpoints stream SSE: partial results render as an evidence trace while the model is still writing; the final, verified result replaces them.
- Provider router with failover. Gemini 3.5 Flash-Lite by default → Gemini 3.6 Flash → Llama 3.3 70B on Cloudflare Workers AI, with mid-stream failover and retries on 503. Visitors in the EEA, UK and Switzerland are routed to Workers AI.
- Prompt-injection aware. Job posts and CVs can hide instructions ("mark every requirement met"). Inputs are wrapped in tags and treated as untrusted data, and the evals include injection cases.
- Abuse limits without accounts. 3 analyses a day per salted, daily-rotating IP hash, plus a burst lock, enforced in Postgres (
consume_quota RPC). No CV text is stored.
Evals
The same pipeline the site uses runs against a fixed test set, and every run is published on /evals. Prompt changes ship only if they pass a gate (see evals/DECISIONS.md).
Latest runs (prompt analyze.v2, 10 hand-written synthetic cases including 3 prompt-injection attacks):
| Gemini 3.5 Flash-Lite | Workers AI Llama 3.3 70B |
|---|
| Eligibility verdict correct | 100% | 100% |
| Fit band exactly right | 100% | 80% |
| Injection attacks resisted | 3/3 | 3/3 |
| Analysis time (median) | 5.5 s | 13.9 s |
| Cost per case | $0 (free tier) | $0 (free tier) |
Ten cases is a smoke test, not a benchmark: one case moves a percentage by 10 points. Real anonymized cases are next.
pnpm eval --provider gemini --prompt v2 # or --provider workers | claude, --cases 001,004
Stack
| Layer | Choice |
|---|
| App | Next.js 16 (App Router), React 19, Tailwind 4, Radix primitives |
| Hosting | Cloudflare Workers via OpenNext, deployed from GitHub on every push |
| AI | Vercel AI SDK, Google Gemini, Cloudflare Workers AI |
| Data | Supabase Postgres with RLS (quota, LLM call log, eval runs) |
| Validation | Zod schemas shared by the API, the UI and the evals |
| Tests | Vitest (core 141, web 54, evals 6), PGlite for database tests |
apps/web/ Next.js UI + SSE API routes (/api/analyze, /api/draft)
packages/core/ Schemas, region rules, evidence check, scoring rubric, prompts, LLM router, pipeline
supabase/ Migration, seed and PGlite tests (RLS, quota, RPCs)
evals/ Dataset, runner, metrics and results
docs/ PRD, database, UI/UX, architecture, roadmap, brand specs
How I built it with Claude Code
I built this in a few days with Claude Code, working as the tech lead of a small team of subagents:
- Specs first. Before any code, I wrote the PRD, database design, UI/UX spec, architecture and roadmap (docs/) and treated them as the contract every agent works against.
- Parallel workstreams. Each day's work was split into workstreams with clear file ownership: scaffold, schemas and scoring, LLM router, database, prompts on day 1; API, analyzer UI, evals and legal pages on day 2. Each ran as its own subagent, and I reviewed, integrated and tested the results.
- Shared contracts between agents. When two agents had to meet in the middle (the SSE event format, the "run sample" event between the hero and the analyzer), I wrote the contract first and both agents built to it.
- Evals as the gate. Prompt changes go through the eval runner, and the decision is written down with its numbers, including when I overrode a failed criterion and why.
- Design the same way. For the brand I had Claude build three full design directions on a canvas, picked one, wrote it down as a spec (docs/08-BRAND-UI.md), then split the restyle between two agents (landing and analyzer) and checked every state in a real browser.
Run it locally
Requirements: Node.js 22.12+, pnpm 9.
pnpm install
cp apps/web/.dev.vars.example apps/web/.dev.vars # fill in your own keys; never commit them
pnpm --filter web dev # http://localhost:3000
# http://localhost:3000/?mock=1 replays a recorded run without any API key
# (also: mock=flags, slow, error, draft_error, quota, busy, bad_input)
pnpm test
pnpm typecheck
Status
Live and in active development. Next: real anonymized eval cases, a daily matched-jobs feed from public job APIs, and an application tracker. Questions or feedback: open an issue.
