Trades Voice-to-Quote AI Automation
Independent AI automation prototype
- My role
- AI Automation & Full-Stack Developer
- Team
- Solo project
- Tools
- Python
- FastAPI
- Whisper
- Ollama
- Qwen
- n8n
- Next.js
- TypeScript
- Supabase
- PostgreSQL
- Gmail
- ngrok
- Timeline
2026 — Prototype
- Description
An AI-assisted workflow that converts spoken trade enquiries into structured job information. A customer's voice message is transcribed locally, analysed by a local language model, converted into structured fields, stored in a database, and surfaced through a review interface so repetitive enquiry processing can be handled faster while keeping a person in control of pricing and communication.
- Context
I built this project to explore a practical use of local AI in a business workflow rather than creating another general-purpose chatbot. Trade businesses often receive job enquiries through informal channels such as phone calls or voice messages. Someone then has to manually extract the customer's details, understand the job, determine its urgency, identify missing information, and decide what happens next. The goal was to automate the repetitive interpretation and data-entry portion of that workflow without pretending that an AI system should autonomously make every pricing or business decision.
Click around...
Challenge
Voice enquiries are unstructured. Customers may describe several issues at once, leave out critical information, speak casually, or provide details in an unpredictable order.
A useful automation therefore cannot simply transcribe audio. It needs to convert that transcription into reliable structured data that downstream systems can use.
The second challenge was deciding where automation should stop. Automatically generating a confident quote from incomplete information would make the workflow less trustworthy rather than more useful.
Constraints
I wanted the core AI processing to run locally where practical, which meant working with local transcription and language models rather than relying entirely on hosted APIs.
Local inference introduces its own limitations around model size, latency, hardware, and structured-output reliability.
The prototype also needed to expose intermediate results so a human could review what the AI extracted rather than hiding the entire process behind an opaque automated decision.
Research
I explored how speech-to-text, local language models, structured outputs, workflow orchestration, and human review could be combined into a practical business pipeline.
Whisper was used to convert incoming audio into text. A locally hosted Qwen model through Ollama then interpreted the transcription and extracted structured job information.
I deliberately separated transcription, interpretation, orchestration, persistence, and review into different parts of the system so each stage could be inspected independently.
Iterations
The first important step was getting reliable local transcription.
From there, I introduced structured extraction rather than treating the model's response as free-form text. The workflow was then connected to n8n so downstream steps such as storage and communication could be orchestrated visually.
Supabase provided persistent storage, while a Next.js interface gave the business a place to inspect the resulting enquiries rather than interacting directly with workflow infrastructure.
The prototype ultimately became an end-to-end demonstration of voice input moving through AI processing into a usable business record.
Key Features
- Voice enquiry ingestion
- Local speech-to-text using Whisper
- Local LLM inference through Ollama
- Structured extraction using Qwen
- Job-detail identification
- Urgency extraction
- Missing-information identification
- Workflow orchestration through n8n
- Supabase/PostgreSQL persistence
- Next.js review dashboard
- Follow-up communication workflow
- Human review before consequential decisions
Final Deliverable
The final prototype demonstrated an end-to-end workflow from raw voice enquiry to structured, reviewable job information.
Rather than using AI as a standalone interface, the project embedded AI inside a larger operational process involving transcription, structured extraction, database storage, workflow automation, and human review.
Takeaway
The main lesson from this project was that useful business AI is often less about generating more text and more about turning messy information into dependable structure.
It also reinforced the importance of separating AI suggestions from business authority. The system becomes more practical when the model handles repetitive interpretation while humans retain control over pricing, exceptions, and customer-facing decisions.