Jev AI for Customer Feedback Analysis

Sep 19

Jev is an AI tool designed to classify customer feedback into categories like bugs or feature requests and identify duplicate issues, returning typed answers with calibrated probabilities instead of generating text.

SignalWhat I tested The system turns raw feedback into ranked lists of top bugs and top feature requests for product managers.

Pattern↑ Jev AI decision engine

I tested Jev @typesafeai inside my AI customer feedback system, on real data: 1,040 public GitHub issues from the Claude Code repo. Anthropic Compared with Claude Sonnet 4.6 on the same calls: Jev was 98% cheaper AND 84% faster while being MORE accurate. Game changer!! It is almost instantaneous. What I tested The system turns raw feedback into ranked lists of top bugs and top feature requests for product managers. Two of its steps don't write anything. They judge: • Is this a bug, a feature request, or spam? • Is this the same problem as one we've already seen? That's what Jev is built for. It doesn't generate text. It returns typed answers with calibrated probabilities. Results on those two steps • Cost: $9.25 → $0.20 (98% cheaper, 47x) • Speed: 3.0s → 0.49s per decision (84% faster) • Accuracy: 96.3% vs 94.5% on bug-vs-feature, scored against GitHub's own labels (946 issues) • Matching duplicates: on par with Sonnet • The whole experiment: 6,240 API calls for $0.53 For a million datapoints, that is about $10,000 in saving. What it doesn't do Jev can't write. Naming and describing each new cluster still runs on Claude, and that was over half my bill, so the pipeline as built drops about 38% overall. The bigger win comes from redesigning around cheap judgment. That's part 2. Takeaway: count how many of your LLM calls are choosing vs. writing. Mine were 2 out of 3.

Original post