TicketForge

Turn meeting chaos into Linear tickets — in 10 seconds. AI extracts, you review, one click pushes to Linear, GitHub, or Notion.

Comment

Inspiration

Every engineering manager I know has the same Friday afternoon ritual: open the meeting transcript, copy-paste fragments into Linear, fix the assignee names the AI got wrong, set priorities by hand, and lose an hour. For a 20-person team, this adds up to $130K/year in lost capacity. I wanted to fix it.

What it does

TicketForge converts meeting transcripts into properly-structured Linear tickets in 10 seconds.

You can paste a transcript, upload a .vtt file, or record live in the browser. The AI extracts action items with:

  • Confidence scores so you review what needs review, not everything
  • Source quotes — every ticket links back to the exact line it came from
  • Smart assignee matching that handles first names, nicknames, and typos
  • Priority and due date inference from natural language

You review in a clean human-in-the-loop UI, then push to Linear, GitHub Issues, or Notion. There's also a Slack bot so your team can trigger extraction directly from any channel.

How I built it

  • Frontend: Next.js 16 (Pages Router) + TypeScript + Tailwind CSS v4
  • AI: Groq (Llama 3.3 70B for extraction + time estimation, Whisper Large V3 Turbo for transcription) with Zod schema validation and auto-retry on invalid JSON
  • Integrations: Linear SDK, GitHub REST API, Notion API, Slack Web API with HMAC signature verification
  • Hosting: AWS EC2 (t3.small), Nginx reverse proxy, PM2, Let's Encrypt SSL
  • Custom domain: ticketforge.aryangorde.com

Challenges I ran into

LLM output reliability. Groq's JSON mode guarantees valid JSON syntax but not schema adherence. Solved with Zod validation + automatic retry with a stricter instruction on the second attempt.

Assignee matching accuracy. Pure Levenshtein matched "Sara" to "Sarah" but also matched "Marcus" to "Marc" at the same score. Built a hybrid scorer with first-name boost and substring containment that handles transcript-style names correctly ~95% of the time.

State between pages. No database, no auth — used sessionStorage to pass extraction results from /extract → /review. Survives refresh, doesn't leak across sessions.

Accomplishments I'm proud of

  • Full end-to-end flow live on a custom domain with HTTPS in 4 days
  • Voice → tickets pipeline that actually feels magical when you demo it
  • Confidence scoring that makes the human-in-the-loop UX feel intentional, not like a workaround for unreliable AI
  • AI-generated time estimate that shows exactly how many minutes the manual work would have taken

What I learned

  • Whisper-Large-V3-Turbo via Groq is fast enough for live demos — sub-2-second transcription on a 30-second clip
  • Slack's Block Kit is genuinely good for non-trivial bot UIs
  • Keeping state in sessionStorage (no DB, no auth) is a legitimate architecture for hackathon tools

What's next for TicketForge

  • Browser extension for one-click extraction from any Zoom/Meet/Otter page
  • Calendar integration to auto-fetch transcripts when meetings end
  • Recurring action item tracking — "Sarah, you said you'd ship the auth migration last week"
  • Self-hostable Docker image for teams that can't send transcripts to a third party

Built With

Share this project:

Updates