SOLV.AI: CUTTING VOICE AI INFERENCE COST BY 99.9% WITH ONNX ENSEMBLES
The brief: build an AI voice complaint management system for an FMCG brand that categorises, analyses, and responds to customer calls automatically. The first prototype used GPT-3.5 end-to-end and cost $1,500 per million tokens. The shipped system costs $1.83. Here's how.
Cost reduction
Latency reduction
Category accuracy
vs $1,500 GPT-3.5
THE ORIGINAL PIPELINE (AND ITS PROBLEM)
GPT-3.5-turbo is excellent at NLU but catastrophically expensive when every customer call goes through it. FMCG complaint volumes are high — thousands of calls daily — meaning the cost model doesn't survive contact with production traffic.
THE ONNX ENSEMBLE APPROACH
WHY ONNX + DISTILBERT INSTEAD OF GPT?
Intent classification is a solved problem for narrow domains. DistilBERT-MNLI in ONNX INT8 achieves 100% accuracy on the FMCG complaint taxonomy (product quality, delivery, billing, returns) at 12ms latency. GPT-3.5 adds latency, cost, and non-determinism to a task that doesn't require it.
The 6-state FSM ensures predictable state transitions. Agentic loops with LLMs are powerful but unpredictable — for a voice interaction where every second counts, a deterministic FSM with LLM generation only at the response layer gives you the best of both worlds.
THE DUAL-FALLBACK ARCHITECTURE
Every critical path has a fallback. If Groq's API is unavailable, the system routes to a locally-running Ollama instance (Llama-3.1-8B). If ElevenLabs TTS is down, pyttsx3 handles synthesis locally. Zero downtime in production.
Written by Neal Daftary — AI & ML Engineer, Nirma University, Ahmedabad