Fixing broken resume imports (OCR + LLMs for 100% accuracy)

The first problem was the most basic one: Odyseek couldn't reliably read a resume. It used outdated OCR that jumbled sections together, and sorting a candidate's job history by date took a maze of slow, fragile workflows.

My first fix was to pair OCR with an LLM. OCR pulls out the raw text, and the LLM turns it into clean, structured data on import. With that combination, import accuracy reached 100%, and we got rid of the date-sorting workarounds almost overnight.


Auditing every part of a resume with AI

Once resumes were coming in clean, we could start judging how good they were. I built a set of targeted AI checks that review a resume line by line:

  • Voice and pronouns: Do bullets use active voice and avoid personal pronouns?
  • Clarity and impact: Is the wording clear, and do the impact statements land?
  • Balance: Does each role have a sensible number of bullet points?
  • Structure: Are the layout, summary length, and number of skills right?

Rebuilding from Bubble into a TypeScript & React career agent

As these features piled up, the original stack started to hold us back. Odyseek was built on Bubble (no-code) and stitched together with several outside tools, including Vellum for running our AI workflows. I decided to rebuild the entire platform from scratch in TypeScript and React and bring all of our AI workflows into one in-house codebase.

1. OCR + LLM Parser 100% Import Accuracy Line-by-Line Audits 2. TypeScript & React Agent In-House AI Workflows Role Matching + Voice Coach 3. Soft-Skill AI ATS 500-1,000 Resume Batches Scores Grit & Custom Traits

The rebuild also changed how people used the product. Instead of filling out forms, users talked to a simple, smart chatbot:

  • Drop in a resume, get a full breakdown. Upload your resume in chat, and the agent runs every formatting, wording, and bullet-count check automatically.
  • Role suggestions and tailored rewrites. The agent suggests roles that fit your background, then rewrites your resume for the exact role you want.
  • Voice interview and value-story coach. Candidates can practice live mock interviews and rehearse telling their personal value story out loud.

Why I built the interview coach with a voice-to-voice model

That last feature involved an important design decision. I could build the voice agent as a traditional pipeline (speech-to-text → agent loop → text-to-speech), or I could use a direct voice-to-voice model.

The deciding factor was that a mock interview doesn't need any new information mid-call. Everything the agent needs (the interview questions, the target role, and the candidate's resume) is loaded before the call starts. Since there was no reason to run an agent loop during the conversation, I went straight to voice-to-voice, and the experience was much better for it:

  • No middleman latency. Each turn skips transcription, an LLM reasoning loop, and a separate speech synthesis step, so replies come right away.
  • A natural conversation. Practicing your value story feels like a real back-and-forth interview, not like waiting on a laggy bot.

Building an AI ATS that scores soft skills across 500 to 1,000 resumes

While all of this served candidates, we also built for the other side of hiring. I created an AI-powered Applicant Tracking System (ATS) that analyzes batches of 500 to 1,000 resumes at once, helping companies like Mike's Hot Honey hire much faster.

Most ATS tools only look for exact keyword matches. We designed ours to read between the lines and score soft skills, sentiment, and custom hiring criteria:

  • Detecting grit from career trajectory. In one batch, a candidate had spent more than 20 years at Yahoo. A traditional filter might have counted that against him as staying in one place too long. Our agent flagged him as high-grit instead, because it traced how he steadily moved up through different roles and took on bigger responsibilities over those two decades.
  • Custom trait scoring. Hiring teams could type in the traits and criteria they cared about for a role, and the ATS scored every resume on how many of them showed up in the candidate's actual history.