Ticket Intelligence

What does your engineering
history actually say?

task2vec turns years of Jira tickets into a semantic map of expertise — showing how individuals evolve, where the project is heading, and whether the two are moving together.

Open Spring explorer    AI Readiness analysis    AI Work Cockpit    Score a ticket →

The idea

Not all engineering work is equal. Some tickets are repetitive patterns the team has solved dozens of times. Some sit at the edge of what anyone knows. Most are somewhere in between. The problem is that nobody has ever had a reliable way to tell them apart — until now.

task2vec embeds every ticket into a semantic map built from your project's full history. That map makes three things visible at once: which work is safe to hand to an AI agent, which needs an AI draft with a human checking the result, and which problems are genuinely hard — the ones that require deep expertise, careful judgment, and the kind of engagement that actually grows an engineer's skills.

The goal is not to replace engineers. It is the opposite: free them from repetitive work so their time is spent entirely on the problems worth solving. AI handles the routine. Engineers own the frontier.

No manual tagging. No surveys. No process changes. Connect your Jira or GitHub, and the data speaks for itself.


Architecture

1
Ingest tickets
Load resolved tickets from Jira or MongoDB. Each ticket carries a key, summary, description, assignee, and timestamp.
MongoDB · JSONL · Jira REST API
2
Embed ticket text
Each ticket is converted to a 3072-dimensional vector using a text embedding model. Embeddings are cached so large corpora only need processing once.
text-embedding-3-large · local cache
3
Project to 2D with UMAP
PCA reduces the embeddings to 50 dimensions first, then UMAP projects them to 2D. Semantically similar tickets land near each other on the map — organically, without predefined categories.
PCA-50 → UMAP-2 · ~45 s for 70 k tickets
4
Cluster and label themes
K-Means groups tickets into semantic clusters in the original embedding space. Each cluster is labelled automatically by an LLM shown a sample of its tickets.
K-Means (k=32) · GPT-4o-mini labelling
5
Analyse careers and strategy
Each contributor's tickets are split into Early, Middle, and Recent phases. Convex hulls and centroids show the shape of each phase on the map. The project strategy vector is the shift from the earliest centroid to the most recent one.
Phase centroids · cosine alignment · UMAP trajectory
6
Generate narratives
An LLM writes a plain-English work story for each contributor and for the project as a whole, grounded entirely in the structured data — no hallucination, no invented facts.
GPT-4o-mini · structured prompt · cached
7
Serve as self-contained explorer
All pre-computed data is embedded into a single HTML file. No server, no database at runtime. Open in a browser and explore.
Plotly.js · WebGL scattergl · ~16 MB HTML

Findings — 69,000 Spring tickets

We ran task2vec on the full Spring Framework Jira archive: 69,156 tickets spanning 20 years of open-source development. The data reveals three tiers of AI readiness — and a trend that reframes the entire case for AI oversight.

22%
Automate
safe for AI, no human review
30%
Assist
AI draft + human sign-off
48%
Escalate
senior engineer required
The trend — automatable work vs. hard problems over time
2010  19% auto  36% escalate  ← project growing, lots of routine work
2013  21% auto  39% escalate
2016  17% auto  55% escalate  ← maturing, easy work done
2018  15% auto  59% escalate
2020   5% auto  68% escalate  ← only hard problems remain
Key insight

AI automation doesn't shrink the escalate pile over time — it accelerates through the easy work and exposes the hard problems faster. By 2020 nearly 70% of new Spring tickets were architecture-level problems. The triage tool becomes more valuable as adoption grows, not less.

Most automatable work types
  1. Documentation corrections and typo fixes
  2. Dependency and version upgrades
  3. Deprecation API replacements
  4. Codebase consistency, renaming, and formatting
  5. Acceptance test failures — known failure pattern, fix-and-verify
View full analysis
Try it now — free
Does this ticket need a human?

Paste any Jira ticket. We find the most similar Spring Framework tickets in our 61k dataset and look up how they actually resolved — to score yours as Automate, Assist, or Escalate in ~3 seconds.

Score a ticket

Key concepts

🗺️

Ticket landscape

All tickets projected into a shared 2D space. Proximity = semantic similarity. Cluster islands emerge naturally without any manual labelling.

📍

Career phases

Each contributor's work is divided into Early, Middle, and Recent thirds. Phase centroids on the map show where focus lived at each stage.

Project direction

The gold arrow is the project's emergent strategy — the vector from the early centroid of all tickets to the recent centroid. No document defines it.

Alignment score

Cosine similarity between a contributor's personal trajectory and the project direction vector, mapped to a 0–100% alignment score.

📊

Theme river

A stacked bar chart showing which semantic clusters a contributor worked in each year — their thematic fingerprint over time.

📝

Work story

An LLM-generated narrative summarising each contributor's career arc, written from structured data with no invented facts.


Get in touch

Interested in running this analysis on your own data, or want to talk about what task2vec could do for your team? Send us a message.