landing page
job finder
mock interview

CareerGuru AI — Project Story

Inspiration

I've watched brilliant people lose dream jobs not because they lacked the skills — but because they lacked the reps.

Interview prep hasn't changed in 20 years. You read a list of common questions. You practice in front of a mirror. Maybe a friend helps you for 20 minutes before losing interest. Then you walk into a room with a stranger who asks something slightly different — and you freeze.

The real problem isn't knowledge. It's the gap between knowing an answer and delivering it under pressure, to a skeptical human, in real time.

And then there's the resume problem. You spend hours writing it. You apply to 40 jobs. You hear nothing. Not because you're unqualified — because an ATS bot filtered you out at 2am before a human ever saw your name.

Two broken systems. Millions of people stuck between them.

I wanted to build something that attacks both — not with encouragement, but with reality. Something that pushes back the way a real hiring manager would. Something that rewrites your resume the way a recruiter actually reads it.

That's where CareerGuru AI started.

What it does

CareerGuru AI is a full-stack career preparation platform with nine interconnected features, built to take a candidate from zero preparation to interview-ready — end to end.

Interview Simulator Select your role, company type, experience level, and interview type — behavioral, technical, or culture-fit. The AI conducts a realistic mock interview, presenting questions one at a time and scoring every answer across four dimensions: Clarity, Specificity, Confidence signals, and Relevance. Feedback slides up after each answer with a coaching comment and a model answer to compare against.

Rejection Mode™ The feature that doesn't exist anywhere else. Instead of a score after each answer, the AI plays a skeptical interviewer who pushes back first — "That's quite general. What specific data drove that decision?" — exactly the way a real hiring manager would in a room that feels too quiet. Only after the follow-up does it reveal the score and private coaching. It's the closest thing to a real high-stakes interview without the consequences.

Resume Rewriter Upload your PDF resume, paste a job description, and watch the AI extract, analyze, and rewrite every bullet point — stronger action verbs, quantified impact, ATS keyword injection. A keyword match score shows the before-and-after in real time: 34% to 81%. That's the difference between getting filtered by a bot and getting a call.

Job Finder Type any role — “React Native Developer,” “AI Engineer,” “Game Developer” — and the AI scans the web in real time using the Medo Web Search plugin to find every relevant opening across job boards, company career pages, and startup listings. No endless tabs. No duplicate posts. Just live opportunities pulled into one place.

Each result comes enriched with salary estimates, required skills, experience level, remote/on-site status, company insights, and direct application links. The AI also highlights which jobs best match your resume, identifies missing keywords, and surfaces hidden opportunities you’d normally miss buried deep in company sites.

Instead of spending hours searching, filtering, and comparing listings manually, you get a constantly updated pipeline of high-fit jobs in seconds.

Performance Dashboard Every session is tracked. Clarity, Specificity, Confidence, and Relevance scores charted over time. AI-generated insight panels identify your patterns — what you do consistently well, what keeps costing you points, and what to focus on this week. A weak question type heatmap shows exactly where your blind spots live.

Question Bank Describe your role and context. The AI generates 20 hiring-manager-calibrated questions — Behavioral, Technical, Culture Fit, and Case Study — with difficulty ratings. Save favorites, practice individually, or run them through the STAR Coach: a 4-step wizard that walks you through Situation → Task → Action → Result, assembles your full answer, and rates it on all four dimensions.

Live Interview Copilot Running in the background during a real interview. Listens via Web Speech API, detects questions in real time, and surfaces STAR memory trigger prompts on a floating 380px panel — without the interviewer knowing. Tracks filler words live. Has an emergency "Blank? Say this →" phrase for the moments your mind goes white.

Body Language Coach Camera-based real-time analysis using MediaPipe — no video ever leaves the device. Tracks eye contact percentage, head stability, expression warmth, and speaking pace every 2 seconds. Post-session report shows a timeline of where you lost eye contact and the top two moments to fix.

Salary Negotiation Simulator Practice the conversation most people never practice. Choose your AI recruiter personality — Friendly, By the Book, or Hard Negotiator. The Hard Negotiator uses silence, implies the offer might be pulled, and pushes back on every counter. The debrief shows dollar amount gained, tactics detected, and a full annotated replay.

Interview Prep Roadmap Enter your target role, interview date, and available hours per day. The AI builds a day-by-day preparation calendar — daily themes, 2–3 tasks per day linking directly to app features, time estimates. If your scores drop mid-preparation, the roadmap automatically inserts a correction day. Export the whole thing to Google Calendar in one click.

How we built it

CareerGuru AI was built entirely using MeDo — Baidu's conversational full-stack app builder.

Every feature was described in plain English across focused multi-turn sessions. No boilerplate. No build configuration. No 3am debugging spirals from misconfigured environments. Just conversation — and a working application on the other side.

The build followed five core MeDo sessions:

Session 1 — Full app scaffold. Described the complete page hierarchy, navigation structure, dark navy design system, and routing. MeDo generated the React frontend, Node/Express backend, and database schema from that description.

Session 2 — Interview AI logic. Described the scoring rubric across four dimensions, the question generation behavior per role and experience level, and the Rejection Mode dialogue flow — including the push-back response before score reveal. MeDo built the entire multi-turn interview engine from that single session.

Session 3 — Resume Rewriter. Described PDF upload, the Gemini Vision extraction pipeline with PDF.js fallback, the side-by-side diff view, keyword match scoring, and one-click PDF export. MeDo wired all of it together — including the animated loading overlay cycling through progress messages while processing runs.

Session 4 — Live features. Described the Live Interview Copilot floating panel, Web Speech API transcription loop, filler word detection, Body Language Coach MediaPipe integration, and Salary Negotiation recruiter personalities. MeDo built all three as independent modules under the Live & Social navigation dropdown.

Session 5 — Dashboard, Roadmap, and polish. Described the performance tracking schema, AI insight generation, weak question heatmap, shareable result cards at 1200×630px, Google Calendar export, and adaptive roadmap logic. MeDo completed the data layer and all remaining UI states.

Plugins used throughout:

Large Language Model — the core engine powering interviews, scoring, resume rewriting, question generation, STAR coaching, recruiter simulation, roadmap generation, and dashboard insights
PDF plugin — resume upload and text extraction pipeline
OCR — fallback for scanned or image-heavy PDF resumes
Web Search — context-aware question generation and real-time ATS keyword validation
Text to Speech — interview questions delivered as audio for full voice simulation
Web Reader — parse job posting URLs directly from LinkedIn, Greenhouse, or any careers page

Total build time: 48 hours. Estimated equivalent manual development time: 4–5 weeks.

Challenges we ran into

PDF extraction across every resume format. Resumes come in two-column layouts, tables, custom icons, embedded fonts, and scanned images. Getting clean structured text — preserving bullet hierarchy, section headers, and formatting — required building a layered extraction pipeline: Gemini Vision first, PDF.js as fallback, explicit error surfacing if both fail. Silent failures here destroy the entire resume feature, so we made every failure loud and recoverable.

Scoring consistency at scale. An AI that gives wildly different scores to nearly identical answers destroys user trust immediately. Getting the four-dimension rubric to behave consistently — so a 7/10 Specificity answer feels like 7/10 whether the role is software engineering or brand marketing — required careful prompt calibration across dozens of test sessions. The heatmap on the dashboard only means something if the underlying scores are stable.

Rejection Mode tone. Too harsh and users feel attacked, not coached. Too soft and it's indistinguishable from normal feedback mode. The right tone — skeptical, professional, a little impatient, like a real hiring manager on their seventh interview of the day — took about fifteen test conversations before it felt genuinely authentic rather than theatrically mean.

Building 14+ interconnected features. Scope management was a constant battle. Three features got simplified mid-build. The Peer Mock Interview Marketplace XP system was descoped to its core matching and rating flow. We kept asking: what does the judge need to see in three minutes?

Accomplishments that we're proud of

Rejection Mode™ is genuinely novel. No other interview prep tool simulates the push-back moment — which is arguably the single most important moment in any real interview. We built a UX pattern that doesn't exist in any competing product, from a one-sentence description in MeDo. A skeptical, impatient AI that pushes back before scoring — that's not a feature. That's a new category.

The Resume Rewriter's keyword match score — watching it animate from 34% to 81% in real time — is one of those product moments that feels like magic and communicates immediate, tangible value without a single word of explanation. That delta is the difference between getting filtered by a bot at 2am and getting a call.

Job Finder pulls live opportunities from across job boards, company career pages, and startup listings in real time — enriched with salary estimates, required skills, remote status, and direct application links. Instead of spending hours in 12 tabs, candidates get a constantly updated, resume-matched pipeline in seconds. It doesn't just find jobs. It finds your jobs.

Company Brief turns a job description URL into a full hiring intelligence report — culture signals, interview patterns, growth trajectory, red flags, and what that specific team actually values in a candidate. Most people walk into interviews having read the About page. Our users walk in knowing how the last three rounds went.

The Interview Prep Roadmap builds a day-by-day preparation calendar from three inputs: target role, interview date, and hours available per day. Daily themes, feature-linked tasks, time estimates — and if your scores drop mid-preparation, it automatically inserts a correction day. It's not a static checklist. It's a plan that responds to how you're actually doing.

What we learned

MeDo's multi-turn sessions are where the product actually gets built. The first prompt scaffolds the structure. The second, third, and fourth refine the behavior, add edge cases, and layer in the logic that makes a feature actually work. Treating each session as a conversation — not a single instruction — unlocked everything.

Describe the feeling, not just the feature. The prompt that built the best version of Rejection Mode wasn't "add a push-back response before the score" — it was "make the user feel like they're sitting across from a skeptical hiring manager who has interviewed seven people today and isn't impressed yet." Emotional specificity produced better output than technical specification every time.

Nine features is too many for a three-minute demo video. Pick three. The Interview Simulator, Rejection Mode, and Resume Rewriter carry the entire narrative. Everything else is supporting evidence. We learned to build everything but only show what makes someone feel something.

The shareable result card was the last feature we built and the first thing people asked about. Adding social proof and challenge mechanics to a solo practice tool changes the entire product dynamic. It was an afterthought that became a product insight: people practice harder when their score is public.

Client-side AI inference is underrated. The Body Language Coach runs MediaPipe entirely in the browser — zero video transmitted, zero storage, zero privacy concern. That constraint produced a better product than a server-side approach would have, because it removed every barrier to user trust.

What's next for CareerGuru

Voice-first interview mode. Full microphone input with real-time transcription, filler word detection — um, uh, like, you know — and a Fluency dimension score added alongside the existing four. The most realistic interview simulation happens out loud, not in a text box.

Industry-specific question packs. Dedicated sets for FAANG, consulting case interviews, investment banking, medical residency, and law — each with calibrated scoring rubrics and role-authentic Rejection Mode personas that behave the way those specific industries actually interview.

Peer challenge mode. Two users practice the same question simultaneously, compare scores and coaching feedback, and compete on a live leaderboard. Weekly ranked challenges by role. Make interview prep something people actually want to come back to every day.

Interview debrief upload. Paste a recruiter's rejection email or post-interview feedback, and CareerGuru identifies the exact dimension that cost you the offer — then builds a targeted 5-day practice plan around it.

Enterprise version. White-labeled for universities, bootcamps, and corporate L&D teams. Cohort dashboards, instructor views, bulk session reporting, and placement rate tracking.

The vision hasn't changed from the first prompt I typed into MeDo:

The interview room will never be the hardest room you've been in.

CareerGuru AI makes sure of that.