Telecom quality: 55 → 91.5
Thank you to Andrej Karpathy for sharing"autoresearch". https://lnkd.in/e3XRmWgF
Karpathy's autoresearch pattern helped us clean 3.7M keywords
Autoresearch has a simple loop: define a metric, mutate, measure, accept if better, reject if worse. No human in the loop.
We had a taxonomy problem: 121K nodes, 3.7M keywords across 4 industries. 36% of telecom sat in lazy "General" and "Other" buckets.
Not useful for anyone. We borrowed the pattern:
One metric, a quality score penalizing generic categories and noise keywords.
Autonomous mutation: the system finds word patterns in unclassified keywords, proposes rules, and tests them automatically.
Accept/reject gate: Score improves? Keep it. Drops? Auto-rollback. No bad changes stick.
Multi-round iteration each round builds on the last. Round 1 catches broadband, Round 2 catches streaming, Round 3 catches billing.
Results across all industries:
- 4,008 nodes reclassified from generic → specific categories
- 81K noise keywords flagged (airlines, restaurants, car rentals hiding in telecom data)
- Telecom quality: 55 → 91.5
- Zero manual decisions
#autoresearch
#ai
