Inspiration

A huge number of students — especially from families who can't afford paid bootcamps — get stuck learning fundamentals through ad-cluttered sites or behind "pay to unlock" paywalls, just to prepare for technical interviews. We wanted to build something genuinely as good as (or better than) paid platforms like GeeksforGeeks or bootcamp courses, but keep it completely free, completely ad-free, with no predatory unlock patterns. This isn't a side feature of the project — it's the entire reason it exists. Under-delivering on content quality was something we treated as actively harmful to that mission, not just a minor gap to fix later.

What it does

EduAI Placement Mentor is a full placement-prep platform covering DSA, Python, Machine Learning, System Design, Web Development, and Software Development Engineering fundamentals, with:

• Structured roadmaps and topic content per domain, including practice problems, resources, and interview-prep material.
• AI-Powered Mock Interviews with four distinct modes:
  • Coding — original, AI-generated coding questions per domain
  • HR/Behavioral — curated question bank with AI fallback once exhausted, graded on communication and STAR-method use, not code
  • Technical Panel — in-depth senior-engineer-style questions (system design, trade-offs, engineering judgment)
  • Mixed — a blend across all three

Each answer gets real-time AI grading with a score, strengths, weaknesses, and a model-answer hint, plus an "Explain More" option for a deeper walkthrough of the ideal answer. Students can also end an interview early and still get scored on whatever they've answered so far, instead of being forced to finish a fixed number of questions.

• AI Resume Generator — takes a candidate's real skills, projects, education, and achievements and produces a clean, ATS-friendly resume in Markdown, explicitly instructed never to invent facts, companies, or numbers the candidate didn't provide.
• Progress tracking, activity heatmaps, and analytics across topics and domains.

Tech stack

• Backend: Python (FastAPI), MongoDB
• AI: Groq-hosted LLMs, with a two-tier model strategy — a fast primary model (Llama 3.3 70B) for routine grading and question generation, with automatic escalation to a larger fallback model (GPT-OSS 120B) for technical-panel mode, longer/more detailed answers, in-depth explanations, and as an automatic retry if the primary model's structured JSON response fails to parse
• Frontend: vanilla JS, with marked.js for rendering AI-generated Markdown output (explanations, resumes) as properly formatted HTML

How we built it

We iterated feature-by-feature, always verifying against the real, running application rather than trusting documentation or assumptions — reading actual source files before building on top of them, querying the live database directly to confirm what content actually existed, and testing every new feature live in the browser before considering it done. The Mock Interview feature, for example, went through an explicit UX pass: it started buried inside a single domain's roadmap page, and was rebuilt into its own standalone, discoverable section with a clean domain-picker → mode-picker → interview flow, after early user testing surfaced that the original placement made the feature effectively invisible.

Challenges we ran into

• Verifying AI behavior instead of assuming it. It's easy to write fallback/escalation logic and assume it works because the code reads correctly — we made a point of testing it live with real answers of varying length and complexity to confirm the larger model was genuinely being invoked when it should be, not just trusting the code path on paper.
• UI state bugs that don't throw errors. A CSS specificity collision once caused every interview panel to render simultaneously instead of one at a time, even though the JavaScript toggle logic was completely correct — a good reminder that "the code looks right" isn't the same as "the app behaves right."
• Rendering AI output safely. LLM responses came back as real Markdown, but early versions displayed it as raw text with literal ** and ## characters — fixed by properly parsing Markdown into HTML rather than displaying it as plain text.

Accomplishments that we're proud of

Building a genuinely free, ad-free, full-featured interview prep and resume tool with real adaptive AI grading — not a static question bank — that scales its own model usage intelligently (cheap/fast model for routine work, escalating to a more capable model automatically when the task actually calls for it) rather than over- or under-using compute.

What we learned

That shipping something meant for real students requires treating "it's technically working" and "it actually works well for the person using it" as two different bars — and holding out for the second one, even under time pressure.

What's next for EduAI Placement Mentor

Expanding domain_details coverage to the remaining planned domains (so Mock Interview isn't limited to just a couple of tracks), deepening existing thin content in Machine Learning and System Design, and adding a proper downloadable/formatted resume export.

Built With

• 3.3
• css3
• fastapi
• gpt-oss-120b
• groq
• html5
• javascript
• llama
• marked.js
• mongodb
• motor
• pydantic
• python
• uvicorn