NeuralCV

NeuralCV AI auto-optimizer upgrade popup
NeuralCV 7-agent pipeline running screen
NeuralCV resume template selection gallery
NeuralCV sign-in and guest access page
NeuralCV before-and-after resume transformation
NeuralCV UX designer resume landing page-Dark Theme
NeuralCV what-if score simulator
NeuralCV UX designer resume landing page
NeuralCV resume builder dashboard preview
NeuralCV Keywords to be Added
NeuralCV Course Suggestions.
NeuralCV resume checker with sign-in prompt
NeuralCV AI cover letter generator
NeuralCV resume strengths and weaknesses
NeuralCV LinkedIn profile optimizer dashboard
NeuralCV ATS risk analysis dashboard

Inspiration

My friend spent three months applying to jobs. Same qualifications as people who got interviews. Same universities, same years of experience. She got nothing back.

I watched her rewrite the same resume twelve times and still not know why it was getting rejected. Then I found out about ATS systems. Automated bots that scan resumes before any human sees them, filtering out candidates based on keyword matching and formatting rules most people have never heard of.

That felt deeply unfair. The problem was not her skills. It was the system. NeuralCV is my attempt to level that playing field.

What It Does

NeuralCV is a full-stack AI platform that analyses any resume against any job description and fixes it in under 15 seconds.

You upload your resume (PDF or DOCX), paste a job description, and six AI agents run in parallel:

1. ATS Score Engine Gives you a 0-100 ATS compatibility score with a grade, a radar chart across six dimensions, and an exact list of what is dragging your score down.

2. Keyword Gap Heatmap Shows you every keyword the recruiter's ATS filter is scanning for, colour coded by whether you have it or not. No guessing. No vague advice.

3. GPT-4o Full Rewrite Rewrites your entire resume with strong action verbs, quantified impact, and the missing keywords injected naturally. You see a live before-and-after diff so you know exactly what changed and why.

4. Cover Letter Generator One click generates a tailored cover letter matched to the specific job. Not a template. An actual letter written around your resume and their JD.

5. LinkedIn Profile Optimiser Rewrites your LinkedIn summary, headline, and experience section to match the target role. Most people ignore LinkedIn optimisation. It costs them.

6. Interview Prep Pack Detects the gaps between your resume and the role, then generates the questions recruiters are most likely to ask about those gaps. Comes with model answer frameworks so you actually know what to say.

All six run on Featherless AI using Llama 3.3-70B (smart agents) and Llama 3.1-8B (fast agents) through a fully OpenAI-compatible pipeline.

How We Built It

The core stack is Next.js 16 with the App Router, React 19, and TypeScript. Every AI feature runs through a dedicated API route that calls Featherless AI.

The landing page uses a hand-written HTML5 Canvas 2D animation that simulates a live ATS scan - drawing skill bars, calculating a score in real time, and rendering keyword tags frame by frame. All DPR-scaled so it stays sharp on retina screens.

For document parsing, unpdf handles PDFs and mammoth handles DOCX files. Both extract clean plain text that gets fed into the agent prompts.

The agent architecture runs five parallel calls on analysis:

Agent 1: Resume structure parser (fast model)
Agent 2: JD requirement extractor (fast model)
Agent 3: ATS gap scorer (smart model)
Agent 4: Career coach - rewrites, interview prep, action plan (smart model)
Agent 5: Counterfactual simulator - what-if skill impact (fast model)

The dashboard renders results across a RadarChart, KeywordVisualizer, RewriteDiff, CareerChatbot, InterviewPrep, CoverLetterGenerator, and LinkedInOptimizer - all custom components built from scratch.

Deployed to Vercel. Zero config, zero cold starts on the static pages.

Challenges We Ran Into

Latency. Running five AI agents in sequence would take 45+ seconds. I restructured the pipeline to fire Agents 1 and 2 in parallel, wait for both, then fire Agents 3, 4, and 5 together. Got total time down to under 15 seconds.

Canvas resolution. The ATS animation was blurry on high-DPI screens because the canvas physical pixel size did not match the CSS display size. Fixed by reading devicePixelRatio at runtime and scaling the canvas context accordingly.

JSON parsing failures. LLMs sometimes wrap JSON in markdown code fences or add trailing comments. Wrote a robust extractJSON function that strips fences, attempts direct parse, then falls back to regex extraction of the first valid JSON block.

Resume parsing edge cases. Some PDFs are image-based with no embedded text. Added a clear error state that asks the user to paste their text directly instead of failing silently.

Accomplishments We Are Proud Of

Six fully functional AI tools in one platform, all running in production
A custom canvas animation that actually looks like a real ATS scan
Before-and-after resume diff that shows tracked changes at the bullet level
A counterfactual simulator that answers "what if I had listed Kubernetes?" with a live score prediction
Switching the entire AI backend from Groq to Featherless AI during the hackathon week and having everything work immediately - same API shape, better model availability

What We Learned

Featherless AI's OpenAI-compatible endpoint made the provider swap trivial. Swapping base URL and model names was all it took. That portability matters a lot when you are building fast.

Prompt engineering for structured JSON output is harder than it looks. Small wording changes in the system prompt can cause the model to add explanations outside the JSON block, breaking the parser downstream. Being very explicit about output format in the system prompt fixed this.

The canvas ATS animation taught me how much polish matters for perceived quality. The actual analysis takes 12 seconds but the animation makes it feel intentional rather than slow.