WarmIntro

Outreach
CRM

Inspiration

In 2026, over 70% of jobs are filled through referrals and warm introductions. Every student knows networking matters — but nobody has a system for it. Students juggle LinkedIn, Handshake, email, and spreadsheets. They blast cold applications with a 2–5% response rate while the highest-ROI activity — building real relationships — gets neglected because it feels ambiguous and overwhelming.

We asked: what if recruiting worked like a sales pipeline? What if you could see the exact math — "I need 100 outreaches to land 1 offer, and I've done 34" — and have AI handle the hardest part: writing genuine, personalized messages that reference your actual shared background with each contact?

What it does

WarmIntro is an AI-powered recruiting relationship engine built around one core insight: networking is a numbers game with knowable conversion rates.

Here's the flow:

Upload your resume (PDF or paste text) → Claude parses it into a structured profile with skills, experience, and target roles
Pick your target industry ("I want car manufacturing") → 50+ companies surface, ranked by how many alumni from your university work there, each with a warmth score
Select your favorites → a funnel dashboard appears showing the pipeline math:

$$100 \xrightarrow{30\%} 30 \xrightarrow{50\%} 15 \xrightarrow{40\%} 6 \xrightarrow{50\%} 3 \xrightarrow{33\%} 1 \text{ offer}$$

Explore alumni connections → click any company to see real LinkedIn profiles of alumni from your school. Each connection shows a "warm path" narrative: "Sarah Chen, MechE '19, was in Rice Robotics Club with you. She's now a Senior Manufacturing Engineer at Tesla."
Draft personalized outreach → Claude generates an email and LinkedIn message referencing your specific shared background — not a template. Edit in-app, copy with one click, or open LinkedIn directly.
Generate follow-ups → Claude writes contextual follow-up messages that add value, not just "bumping this."
Track everything → log replies, book coffees, earn referrals. Each action advances your funnel, earns XP, and triggers AI coaching tips. A CRM Kanban board tracks every connection across pipeline stages.
Stay motivated → XP system, levels ("Networking Novice" → "Warm Intro Master"), 7-day streaks, achievement badges with confetti, and an opt-in leaderboard.

The key differentiator is the funnel dashboard — students always know exactly where they stand, what their conversion rates look like, and how many more actions they need. It removes the ambiguity that paralyzes most job seekers and turns networking from a dreaded chore into a game with clear progress indicators.

How we built it

Two developers. Four hours. 6,047 lines of TypeScript. Zero merge conflicts.

We used Claude Code as our primary development tool, with a carefully designed parallel workflow:

CLAUDE.md as AI team coordination. We created layered CLAUDE.md memory files — a root file with the shared product vision and API contract, plus directory-specific files with each developer's instructions. Claude Code automatically loaded the right context for each developer. Dev A's Claude knew backend prompts and scoring algorithms; Dev B's Claude knew component build order and Tailwind patterns. Neither could step on the other's territory.

Architecture for parallel development. We split the codebase by directory ownership. Dev A (backend) owned API routes, services, and data. Dev B (frontend) owned pages, components, and hooks. We defined a shared TypeScript type contract in the first 15 minutes and froze it — 5 API endpoints with strict request/response interfaces.

Mock-first frontend. Dev B started building all UI with full mock data baked into a useApi.ts hook. When Dev A's real endpoints came online, we flipped a single USE_MOCKS flag to false. The frontend was never blocked by the backend.

Claude is the intelligence layer, not a chatbot. Claude powers 7 distinct features — each requiring genuine reasoning, not templates:

Resume parsing (unstructured PDF/text → structured profile)
Company-alumni matching with warmth scoring
Real LinkedIn profile discovery via Serper API
Warm path narrative generation (finding shared background)
Personalized outreach drafting (email + LinkedIn)
Contextual follow-up generation
AI coaching tips after each pipeline stage advance

Challenges we ran into

Alumni stability was the hardest problem. Serper API returns different LinkedIn profiles on each call, so navigating away from a company page and returning would shuffle the alumni list — breaking the "mark as sent" state and double-awarding XP. We solved this with server-side response caching by companyId plus client-side deduplication using both a sentAlumniIds Set on the server and React Context tracking on the client.

Prompt engineering for genuine outreach. Getting Claude to write messages that sound like a real student — not a networking robot — took multiple iterations. The breakthrough was providing rich context about shared background (clubs, professors, overlapping years) and explicitly banning phrases like "pick your brain" and "networking opportunity."

Coordinating two Claude Code sessions. Making two AI-assisted developers work on the same repo without conflicts required upfront architectural discipline. The CLAUDE.md layering system was our key innovation — each dev's AI stayed in its lane because the memory files were scoped by directory, not by manual switching.

Accomplishments we're proud of

6,047 lines of working TypeScript across 41 files and 13 commits in 4 hours — with zero merge conflicts
The funnel dashboard — "34/100 outreaches, on track for 1 offer" makes networking feel concrete
Real alumni discovery — not mock data, actual LinkedIn profiles matched to your university
A full CRM with Kanban board, stage tracking, and AI-analyzed notes — in a hackathon project
Zero UI libraries — pure Tailwind CSS for the entire frontend. Lean stack, 5 runtime dependencies
All Claude AI features running live — no mock fallbacks, every feature hits the real API

What we learned

CLAUDE.md files are incredibly powerful for team coordination. Defining dev boundaries in markdown that AI reads automatically is a workflow pattern we'll use in every future project.
Mock-first frontend development eliminates the #1 hackathon bottleneck (frontend waiting on backend).
The warmth scoring algorithm (same major +20, overlapping years +15, shared clubs +10) produces surprisingly good connection rankings — simple heuristics + Claude reasoning > complex ML.
Gamification isn't a gimmick. The XP and streak system genuinely makes networking feel less overwhelming. We felt this even during our own testing.

What's next

University partnerships — integrate with career services offices and official alumni databases for verified connections
Real Gmail OAuth — move from copy-paste to one-click send with open/reply tracking
Voice interview prep — when a contact reaches interview stage, Claude runs a mock interview with real-time feedback
Cross-university networks — expand beyond Rice to any university, with alumni data from LinkedIn partnerships
Analytics dashboard — conversion rate insights across industries, roles, and outreach styles
Mobile app — push notifications for follow-up reminders and streak maintenance