I spent 4 months job searching with no callbacks. ChatGPT kept inventing stuff I never did. Teal and Rezi made me edit a base resume per job, which is what I was trying to skip. I built PatchWork instead. It takes everything you've ever written about your career, merges it into one profile, and generates a tailored resume per job description.
The hard part was deduplication. The same accomplishment shows up across five resumes worded five different ways, and different metrics survive in different versions. If you just pick one as the winner, you throw away real information. Designed a 5-pass deduplication pipeline, combining deterministic code with AI to get it pretty close to right.
A few of the messier cases:
Umbrella entries. Some resumes list "2018 to 2022, Company X" as one block when it actually covers three different roles. The bullets have to get redistributed to the right titles before anything else happens.
Bullet bleed. The same accomplishment ends up listed under different roles across different documents. Detecting it cleanly is harder than the obvious approach makes it look, because near-duplicates with different metrics are sometimes the same thing and sometimes genuinely two different wins.
Lost context. One version of a bullet has the dollar figure, another has the deadline, a third has the team size. Picking one drops the other two. PatchWork synthesizes across the group so the resume generator can pull whichever detail fits the job.
Project vs experience drift. Same product shows up as an "Experience" entry on one resume and a "Project" on another, sometimes under different brand names. They need to collapse into one thing.
Once the profile is built, applying is paste the JD, click, two minutes. The generator scores each role in the profile against the JD, allocates bullets per role, and writes them with JD keywords baked in for ATS matching. There's a hard ban in the prompt against "spearheaded," "pivotal role," and the rest of the AI-resume cadence that recruiters now filter for.
Every bullet in the output traces back to a source document via a clickable pill. When the model proposes a rewrite it's unsure about, the original wording stays in the resume and a flag icon appears next to it. Click the flag and you see the rewrite, the reasoning, and approve or dismiss it.
Result on me: 0 callbacks in 4 months across roughly 40 applications using other tools. 7 callbacks from the next 10 applications using PatchWork resumes I didn't manually edit. N=1, but the change was sharp enough that I'm running with it.
What's still rough:
The mobile experience. No one keeps resume files on their phone and even if they did, uploading them would be a nightmare. Thinking I'll enable Google Drive access auth to mitigate this in future. Might make it desktop only and "email yourself a link" on mobile for now.
Converting: creating a homepage experience that convinces people to try the tool is hard. I know that people would be impressed by it if they tried it, but I'm not sure how to get over that hump.
Being a resume tool. I think everyone looks down on resume tools because a lot of the ones out there don't do a good job and still make you do so much work. This is the foil to those though, so I'm trying to get that message to be clear and believable.
The hard part was deduplication. The same accomplishment shows up across five resumes worded five different ways, and different metrics survive in different versions. If you just pick one as the winner, you throw away real information. Designed a 5-pass deduplication pipeline, combining deterministic code with AI to get it pretty close to right.
A few of the messier cases:
Umbrella entries. Some resumes list "2018 to 2022, Company X" as one block when it actually covers three different roles. The bullets have to get redistributed to the right titles before anything else happens.
Bullet bleed. The same accomplishment ends up listed under different roles across different documents. Detecting it cleanly is harder than the obvious approach makes it look, because near-duplicates with different metrics are sometimes the same thing and sometimes genuinely two different wins.
Lost context. One version of a bullet has the dollar figure, another has the deadline, a third has the team size. Picking one drops the other two. PatchWork synthesizes across the group so the resume generator can pull whichever detail fits the job.
Project vs experience drift. Same product shows up as an "Experience" entry on one resume and a "Project" on another, sometimes under different brand names. They need to collapse into one thing.
Once the profile is built, applying is paste the JD, click, two minutes. The generator scores each role in the profile against the JD, allocates bullets per role, and writes them with JD keywords baked in for ATS matching. There's a hard ban in the prompt against "spearheaded," "pivotal role," and the rest of the AI-resume cadence that recruiters now filter for.
Every bullet in the output traces back to a source document via a clickable pill. When the model proposes a rewrite it's unsure about, the original wording stays in the resume and a flag icon appears next to it. Click the flag and you see the rewrite, the reasoning, and approve or dismiss it.
Result on me: 0 callbacks in 4 months across roughly 40 applications using other tools. 7 callbacks from the next 10 applications using PatchWork resumes I didn't manually edit. N=1, but the change was sharp enough that I'm running with it.
What's still rough:
The mobile experience. No one keeps resume files on their phone and even if they did, uploading them would be a nightmare. Thinking I'll enable Google Drive access auth to mitigate this in future. Might make it desktop only and "email yourself a link" on mobile for now.
Converting: creating a homepage experience that convinces people to try the tool is hard. I know that people would be impressed by it if they tried it, but I'm not sure how to get over that hump.
Being a resume tool. I think everyone looks down on resume tools because a lot of the ones out there don't do a good job and still make you do so much work. This is the foil to those though, so I'm trying to get that message to be clear and believable.