Inspiration

All of us have had the same moment. You pay for an online course on a topic, and two months later your own Semester 2 module teaches it for free. Meanwhile, the skills that job ads actually ask for (shipping code with a team, reviewing a pull request, explaining a design out loud) aren't on any timetable.

The gap is real. In the Institute of Student Employers' 2025 survey of 148 employers, 77% said students with a placement arrive with better skills. Concern about graduates' verbal communication rose from 29% to 42% in a year. 48% worried that AI-written applications misrepresent what students can really do. Students need practice that looks like real work, and proof that can't be faked.

The gap isn't equal either. First-generation, international and commuter students start with the smallest networks. They're the last to hear what industry wants, and the least likely to know someone a year ahead who can show them the way.

We built Keystone for Aisha, a first-year international CS student. She wants to be an ML engineer and doesn't know anyone at Manchester yet.

Your degree is the foundation. Keystone builds the bridge to industry.

What it does

  1. Reads three things. It reads your university's syllabus (Firecrawl), what industry is hiring for right now (live job ads plus years of Hacker News "Who is hiring" history), and what you've actually built (GitHub).
  2. Builds the Bridge Map. This is a knowledge graph where each industry skill is marked proven, done in your degree, coming later in your degree, or a real gap.
  3. Plans a roadmap that builds on your degree. Modules you'll take later count too, so it never tells you to relearn something your degree is about to teach you. Your hours go only on true gaps, and every recommendation comes with a receipt that shows the module, the job-ad trend or the evidence behind it.
  4. Puts you in a study pod with people who are strong where you're stuck, not copies of yourself, and it explains why each person is in the pod. An AI Project Architect creates a real GitHub repo with scoped issues. An AI Tech Lead reviews your pull requests the Socratic way: it asks questions instead of handing over the answer.
  5. Turns merged PRs into proof. A merged pod PR becomes verified skill evidence for the author, and code-review evidence for the reviewer. A proof ladder (try → build → teach) plus micro-challenges feed a skill passport. A momentum score ranks growth, not raw output, so newcomers aren't buried under people who've coded for years.
  6. Shows you how others got there. You can see the paths of seniors who started where you are, ask practitioners for guidance (you send the drafted messages yourself), and practise in voice mock interviews.
  7. Helps lecturers. Lecturer Insights shows anonymous gaps across the cohort and fast-rising skills that the syllabus is missing.

This covers all three of Reply's pillars: learning (roadmap, pods), teaching (lecturer insights) and assessment (proof from evidence, mock interviews, micro-challenges).

How we built it

  • App: Next.js 16, React 19 and TypeScript, with Tailwind and shadcn/ui, TanStack Query, and zod schemas at every boundary. The Bridge Map uses react-force-graph-2d and the charts use Recharts.
  • Agents: An orchestrator coordinates 14 registered agents: Syllabus, Skill Mapper, Market, Trend Forecaster, Evidence, Roadmap, Matchmaker, Project Architect, AI Tech Lead, Trajectories, Guides, Voice and Automations. Each run streams over server-sent events into a live Agent Activity panel, so you can watch them work.
  • AI and data: OpenAI gpt-4.1-mini with structured outputs, a retry that repairs bad JSON, and a response cache. Firecrawl handles scraping, Octokit handles GitHub, and ElevenLabs powers mock interviews and an audio weekly briefing.
  • Automations: six n8n workflows: market refresh, Monday briefing, pod kickoff, pod nudge, opportunity watcher and newcomer welcome.
  • Keystone Coach: a chat agent that plans, calls tools, asks a human before anything with side effects, and checks its answer before sending it. The same typed tool registry is also served as an MCP server, so other AI agents can use Keystone's tools.
  • Shipping: GitLab CI runs a secret scan, typecheck, lint, unit tests, a build, Playwright end-to-end tests and axe accessibility checks, then deploys to Vercel.

Challenges we ran into

  • Five people, six hours, an empty repo. We wrote a shared contract first (types, API shapes and scoring rules) and stubbed every module against it. That let all five of us build in parallel from the start, with each feature on its own branch and git worktree.
  • Live AI on stage is risky. A model call can time out mid-pitch. Every agent has a deterministic fallback, and a DEMO_MODE runs the whole flow from fixtures. The public deploy runs in demo mode with no API keys, and the live AI runs on our own machine.
  • Language models invent numbers. The Coach checks every draft. Each number or percentage in its answer must trace back to the student's data or a tool result, and anything it can't trace is rewritten out.
  • Everything the agents read is untrusted. PR descriptions, scraped pages and help-feed posts can all hide instructions aimed at the AI. We built a prompt-injection shield that catches instruction overrides, attempts to extract secrets, hidden Unicode (zero-width, bidirectional "Trojan Source" and tag characters), image links that leak data, and base64 payloads. It cleans the text before the model sees it. On top of that, the Tech Lead can only comment and never approve or merge, and the model never holds a secret. Red-team evals check all of this in CI.
  • Messy syllabi. University pages come back as markdown tables in many shapes. Pulling out clean module titles, credits and compulsory status needed its own parser fixes.
  • Staying in scope. Reply's challenge excludes admin and networking. So we designed the people features as peer learning: pods are learning by teaching, trajectories are near-peer mentoring, and guides are learning from practitioners.

What we learned

  • Agree the contract first, then write code. Agreeing types and scoring rules up front let us merge dozens of feature branches in a single day without breaking each other's work.
  • Receipts build trust. Students act on a recommendation when they can see what's behind it: this module, these job ads, that pull request.
  • What you measure decides who wins. Switching the leaderboard from output to growth changed who appears on it, and that's who Keystone is for.
  • Agents that read the web need a security model from the first line. It's much harder to add later.

What's next

  • A pilot with one University of Manchester CS module, with Lecturer Insights going to the module lead.
  • More universities' syllabi and more target roles.
  • Skill credentials that employers can verify, and a "Bridge Week" launch campaign to bring the first cohort onto the platform.

Built With

  • axe-core
  • elevenlabs
  • firecrawl
  • github-api
  • gitlab-ci
  • hacker-news-api
  • model-context-protocol
  • n8n
  • next.js
  • node.js
  • octokit
  • openai
  • playwright
  • react
  • react-force-graph
  • recharts
  • server-sent-events
  • shadcn/ui
  • tailwind-css
  • tanstack-query
  • typescript
  • vercel
  • vitest
  • zod
Share this project:

Updates

Submission history