Enhance AI Call Agent NLU bug fixes one hour work
Budget / Salary₹600–1,500
TypeFreelance project
LocationRemote
Posted1 hour ago
Our AI-driven phone agent generally hears callers just fine, yet its replies drift off-topic whenever a conversation becomes more complicated. The root of the problem is response relevance caused by weak natural-language understanding, especially when users ask multi-step or nuanced questions.
I’d like you to audit and refine the NLU pipeline—intent classification, entity extraction, dialogue state tracking, and any contextual memory logic—so the agent can reliably interpret and answer complex queries. You may work with whatever stack you prefer (Rasa, Dialogflow CX, Amazon Lex, custom transformer models, etc.); I simply need measurable improvement in real-world scenarios.
Deliverables
• Updated NLU model(s) and training data
• Revised dialogue flows or intents that demonstrate correct handling of at least 10 diverse, complex caller scenarios
• Brief deployment guide or pull-request notes so my team can roll the upgrade into production
Acceptance criteria
1. For the supplied test set of complex queries, precision and recall of intent recognition exceed current baseline by at least 20 %.
2. Live call simulation shows contextually accurate answers in 9 out of 10 trials.
3. No regression in existing simple-query performance.
If this sounds like your wheelhouse, let’s get started—the sooner callers stop getting confused answers, the better.
I’d like you to audit and refine the NLU pipeline—intent classification, entity extraction, dialogue state tracking, and any contextual memory logic—so the agent can reliably interpret and answer complex queries. You may work with whatever stack you prefer (Rasa, Dialogflow CX, Amazon Lex, custom transformer models, etc.); I simply need measurable improvement in real-world scenarios.
Deliverables
• Updated NLU model(s) and training data
• Revised dialogue flows or intents that demonstrate correct handling of at least 10 diverse, complex caller scenarios
• Brief deployment guide or pull-request notes so my team can roll the upgrade into production
Acceptance criteria
1. For the supplied test set of complex queries, precision and recall of intent recognition exceed current baseline by at least 20 %.
2. Live call simulation shows contextually accurate answers in 9 out of 10 trials.
3. No regression in existing simple-query performance.
If this sounds like your wheelhouse, let’s get started—the sooner callers stop getting confused answers, the better.
Apply on Freelancer →
Project sourced from Freelancer.com. Applications happen directly on the original platform — we never collect your data.