← Selected WorksGitHub

AI / Voice / NLP · 2025

SOLV.ai

AI-powered voice complaint management at $1.83 / million tokens.

SOLV
// Overview

SOLV.ai replaces expensive LLM API calls with a lean on-device ONNX pipeline for FMCG complaint triage. A dual-model ensemble — DistilBERT-MNLI for zero-shot classification and MiniLM-L6 for semantic similarity — runs inside a 6-state FSM that orchestrates 5 specialised agents. Dual LLM fallback (Groq → Ollama) and dual TTS ensure continuity if any single service degrades. The result: 65% latency reduction and 100% category accuracy at a fraction of GPT-3.5 cost.

// Specs

Specifications

DomainFMCG Voice AI
Latency12ms (down from 35ms — 65% reduction)
Accuracy100% category classification
Cost$1.83 / M tokens (vs $1,500 GPT-3.5)
Architecture6-state FSM, 5-agent pipeline
// Features

Features

01

Dual-model ONNX ensemble — DistilBERT-MNLI + MiniLM-L6 for zero-shot classification

02

6-state finite state machine with 5 specialised AI agents for complaint routing

03

Dual LLM fallback (Groq → Ollama) + dual TTS pipeline for high availability

04

65% latency reduction — inference time dropped from 35ms to 12ms

05

100% category accuracy across all tested complaint types

06

Cost: $1.83 / million tokens vs $1,500 / million for GPT-3.5

// Tech

Tech Stack

ONNXDistilBERTFastAPIGroq

Interested in
SOLV.ai?

Explore the source code, architecture notes, and full implementation.

View on GitHubGet in Touch →