Scholarship Assitant

Enternace
Asking user more story that we can use in application
Application essay
Enhanced resume
Out reach email for user

Inspiration:
Most students recycle the same generic essay across dozens of scholarships. Meanwhile, committees quietly optimize for very specific values buried in their mission statements and past winner bios.

We wondered: What if an AI could reverse-engineer what a committee really cares about, then adapt your story to fit—without making anything up?

That idea became ScholarMatch: a “Chameleon Engine” that reshapes real experiences to match the unique personality of each scholarship.

🧠 What it does

ScholarMatch is an agentic workflow that:

Decodes the Scholarship – Scrapes the donor’s site, “About” page, and past winners to build a structured Value Profile.
Profiles the Student – Ingests resumes and past essays into a vector database.
Finds the Gaps – Compares values vs. experiences and, when something’s missing (e.g. “service” or “leadership”), interviews the student for a relevant story.
Drafts a Tailored Essay – Generates a scholarship-specific essay using the organization’s tone and vocabulary.
Automates Outreach – Identifies decision-makers and drafts a warm, personalized inquiry email.

How

⚙️ we built it

We used LangGraph to orchestrate 7 focused agents:

Scout – Scraping (Firecrawl, Tavily) + contact discovery
Profiler – PDF parsing + embeddings with ChromaDB
Decoder – Criteria + tone extraction (Claude 3.5 Sonnet)
Matchmaker – Semantic alignment between values and stories
Interviewer – Human-in-the-loop Q&A for missing info
Optimizer – Reorders resume bullets for this scholarship
Ghostwriter – Final essay + outreach email

Frontend: Next.js + Tailwind CSS (glassmorphism UI).
Backend: FastAPI serving the LangGraph workflow via async endpoints.

🚧 Key challenges

Avoiding “generic essay” mode → solved with an explicit Gap Analysis stage before writing.
State & interruptions → used LangGraph’s interrupt_before to pause for user answers, then resume with new context.
Messy websites → hardened the Scout agent to handle SPAs, PDFs, and legacy HTML.

🏆 What we’re proud of

The Interviewer agent that asks laser-specific follow-ups instead of guessing.
End-to-end flow: from decoding criteria → drafting essays → networking outreach.
A fast RAG pipeline that makes the whole experience feel interactive, not batchy.

What we learned:

Narrow, well-scoped agents beat one “do-everything” model.
Upstream analysis quality (Decoder) determines downstream essay quality.
LangGraph’s cyclic and conditional edges are ideal for human-in-the-loop workflows.

What’s next :