ApplyCove

A job application automation platform that applies on a candidate's behalf while they watch live and can take over the browser mid-run. Built solo end to end (product, backend, frontend, browser extension, scrapers, infra and growth) and taken to 138,222 applications submitted for 2,326 users in under three months.

Open live site

source: not public

138,222
applications submitted: 494,574
questions auto-answered: 728,653
jobs screened out: 11,683
hours of manual work saved: 2,326
registered users: 11.6%
free to paid

Built with

Node.js
Express
PostgreSQL
Prisma
pgvector
Redis
Puppeteer
React 19
Astro
Manifest V3
AWS ECS Fargate
AWS IoT Core
Cloudflare R2
Docker
LLM cascade

Impact

138,222 job applications submitted and 494,574 screening questions auto-answered, with another 728,653 jobs screened out before applying.
About 11,683 hours of manual applying saved, at a sustained 19.46 applications per user-hour, peaking between 09:00 and 11:00 IST.
2,326 registered users in under three months at an 11.6% free-to-paid conversion rate, on zero paid acquisition spend, all organic search and word of mouth.
~90,000 lines across seven apps, five shared packages and a CLI in a pnpm monorepo, shipped solo in six months.

Architecture

Two automation engines: a Puppeteer worker on ECS Fargate for authenticated job boards, and a Manifest V3 extension that runs the apply loop inside the candidate's own browser for five ATS platforms, which removes the anti-detection problem entirely.
Live session streaming: the worker's X display captured with ffmpeg x11grab into one-second HLS segments on Cloudflare R2, progress events over MQTT on AWS IoT Core, and browser takeover through KasmVNC tunnelled over cloudflared.
Express API with 107 handlers behind nginx, Postgres with pgvector and tsvector full-text search, self-hosted Redis, and blue-green deploys driven by GitHub Actions runners the host registers itself.

Engineering highlights

Three-tier answer resolution: deterministic regex matcher, pgvector semantic cache at 0.15 cosine distance, then a four-provider LLM cascade. Most of the 494,574 questions answered never reached a model.
Concurrency pushed into Postgres instead of application locks: partial unique indexes, SELECT ... FOR UPDATE SKIP LOCKED job claiming, compare-and-swap idempotency on every payment settle path, and a PL/pgSQL trigger settling runtime quota against a FIFO credit wallet.
Recommendation engine written from scratch: Postgres full-text retrieval, two in-memory IDF passes, a declarative capped scoring table, company diversity with overflow backfill, all in a Piscina worker pool.
Six ATS scrapers over 16,597 vendored tenant company lists, including a facet-subdivision strategy that works around Workday's hard 2,000-row query cap, feeding hash-gated upserts that write only changed rows.
Every Redis dependency fails open in a named way: rate limiting falls back to a per-worker memory store behind a circuit breaker, caches fall back to the database. Redis going down degrades the product, it does not stop it.

Product and growth

Distribution built as an engineering surface because there was no ad budget: programmatic city and role pages, hub-and-spoke internal linking, and free ATS-checker tools built to make exactly one model call each so they stay cheap to give away.
AI search handled deliberately: RFC 9110 content negotiation serving Markdown twins to AI crawlers, a generated llms.txt, 30 schema.org entity types, and a 40-assertion conformance suite that gates deploys.
Honest instrumentation: a session-outcome classifier reading freeform reason text, because raw status conflated user-stop, quota exhaustion and real crashes and reported 39% aborted when the hard-failure rate is under 5%.
Owned product decisions, visual design, payments across three gateways with region-based routing, transactional email, SEO and lifecycle marketing.