840 applications. 6 months. One AI agent. I built JobApplier.site to automate my own job search, and then I used it — extensively. Here's every number, every insight, and every mistake I made along the way.

The Setup

I'm a senior software engineer with 8 years of experience. Backend-heavy: Python, TypeScript, PostgreSQL, AWS. I was looking for remote senior or staff roles at Series A–D startups and mid-size tech companies. My target compensation was $180–220K base.

The agent I built (which became JobApplier.site) scanned LinkedIn, company career pages, and ATS portals. For every job matching my criteria, it:

Extracted the job description
Calculated a semantic match score (0-100) against my resume
If score ≥ 65: tailored my resume, generated a cover letter, and applied
If score < 65: logged it and moved on

The Raw Numbers

Total applications	840
Average match score	78.4
Callback rate (phone screen)	14.6% (123 callbacks)
Technical interview rate	7.1% (60 interviews)
Final round rate	2.6% (22 final rounds)
Offers	6
Cost (API + compute)	$127.40 total (~$0.15/application)
Time saved (estimated)	280+ hours

What Worked

1. Per-JD Tailoring Was the Biggest Factor

In my first 100 applications, I ran an A/B test: 50 with generic resume, 50 with per-JD tailored resume. Results:

Generic: 3 callbacks (6%)
Tailored: 9 callbacks (18%)

3x improvement. After that, I never sent a generic resume again. Every single application got a tailored resume with JD-specific keyword injection, reordered bullets, and an adjusted title line.

2. Match Score Threshold Was Crucial

Early on, I set the threshold at 50 — basically "apply to anything remotely relevant." My callback rate was 8%. When I raised it to 65, callbacks jumped to 14.6% while total applications dropped by only 15%.

The insight: a higher threshold doesn't mean fewer interviews. It means fewer wasted applications and a higher signal-to-noise ratio.

3. Cover Letters Moved the Needle

I tested applications with and without cover letters for 200 applications:

Without cover letter: 11.2% callback
With tailored cover letter: 15.8% callback

The key word is "tailored." The cover letter had to reference specific things from the JD — a product the company shipped, a technical challenge they mentioned, anything specific. Generic cover letters performed the same as no cover letter.

4. Stealth Automation Mattered

Around application #200, I got a LinkedIn rate limit warning. I had been applying too fast — 30+ applications in a day with consistent timing intervals. After implementing warmup ramps, Gaussian-distributed delays, and daily caps (max 12 applications/day), I never got flagged again.

What Failed

1. Applying to Jobs Below My Level

I applied to some mid-level roles thinking I'd be a shoo-in. Callback rate on mid-level roles: 8%. On senior roles: 16%. Recruiters are pattern-matching level, and an overqualified candidate is a flight risk. I stopped applying below-level at application #300.

2. Custom Question Hallucinations

In 12 applications, GPT answered a custom question about visa sponsorship differently than my configured answer. This is why I rebuilt the form handler to use deterministic answers for all critical questions (authorization, sponsorship, salary, availability) and only GPT for genuinely freeform questions.

3. Weekend Applications Underperformed

Applications submitted Saturday-Sunday had a 9% callback rate vs. 16% for Tuesday-Thursday. Either recruiters buried in Monday email triage skip weekend submissions, or ATS sorting pushes recent applications to the top and Monday/Tuesday applications leapfrog weekend ones.

I configured the agent to only apply Tuesday through Thursday.

4. Some ATS Portals Were Brittle

Workday and iCIMS forms broke regularly — login flows changed, form fields moved, CAPTCHA challenges appeared. Greenhouse, Lever, and Ashby were far more stable. I ended up deprioritizing Workday companies and focusing on Greenhouse/Lever shops.

The Money Math

Let me put this in perspective:

840 applications × 20 minutes each = 280 hours of manual work saved
Total cost: $127.40 in API calls (GPT-5.4 for tailoring + cover letters)
Cost per callback: $1.04
Cost per offer: $21.23

I spent $127 and saved 280 hours. If I value my time at $100/hour (conservative for a senior engineer), that's a 220x ROI.

The Unexpected Lessons

The job search is a funnel, not a lottery. Optimize conversion rates at each stage, not volume at the top.
Timing matters more than you'd think. Tuesday morning applications outperform Friday afternoon by 2x.
Deterministic > probabilistic for critical fields. Never let AI guess on visa sponsorship, work authorization, or salary requirements.
Match score thresholds save you from yourself. Without a floor, you'll apply to everything. The floor forces quality.
Six offers from 840 applications is actually a great output. Industry averages suggest 200+ applications per offer for senior roles. I was at 140 per offer — and I wasn't manually doing any of the work.

What I'd Do Differently

If I were starting over:

Start with a match threshold of 70, not 50
Only apply Tuesday through Thursday from the beginning
Spend more time on my master resume YAML before starting the agent
Set up interview prep documents from day 1 (I added this feature at application #400)

Try It Yourself

Everything I described is now available as JobApplier.site. The free tier gives you 3 AI-tailored applications per month — enough to see if the tailoring approach works for your profile before committing.

Or if you're technically inclined, the full agent is open-source. Fork it, audit it, run it yourself.

Either way: stop copy-pasting. Start tailoring. The data speaks for itself.

I Applied to 840 Jobs with AI — Here's What I Learned