The problem
Field reps needed account insights between customer visits without working through dashboards.
← Selected work
Voice AI for B2B sales
Voice Sales Assistant
QPOI designed and built a voice and chat assistant around a wholesale distribution software provider's existing sales analytics. It lets field reps ask about their accounts in everyday language, connects to the client's user accounts and runs on its infrastructure.
- Client
- Confidential
- Sector
- Distribution
- Platform
- Web PWA + voice
- Stack
- React + NestJS
- Engagement
- Full product build
- Status
- In client testing
The project at a glance
What QPOI delivered
QPOI took the assistant from discovery and architecture through implementation and deployment, integrating existing sales analytics and user accounts.
Delivery evidence
Deployed on the client's infrastructure and refined through repeated testing with its staff. Supported by more than 2,400 automated tests and live-model evaluations.
Question-to-answer pipeline, as builtView diagramHide diagram
1.0
The challenge
The client's platform already held account insights: declining sales, margins against benchmarks and products a customer had stopped buying. Accessing those insights meant reading dashboards, which did not suit field reps moving between customer visits.
The client wanted reps to ask questions aloud and receive answers about accounts they were permitted to access. QPOI turned that requirement into a working product, from technical discovery and architecture through implementation, deployment and repeated testing with the client's staff.
2.0
Core features
2.1
Real-time voice
Speech-to-speech conversation with barge-in: the rep can talk over an answer or tell it to stop, like a real conversation. A typed chat interface shares the same pipeline.
2.2
Approved database queries
The model never writes SQL. It selects from a set of approved, parameterised query tools, and the server owns every query that touches the database.
2.3
Grounded answers
Account figures come from the data. Most on-screen account answers use templates; where a model writes the wording, a failed figure check triggers a template fallback. Spoken figures are checked after the answer is spoken, with mismatches monitored and alerted on.
2.4
Permission-scoped data
Each rep sees only their own accounts. Scope comes from the client's own user records and is enforced on the server for every request, and restricted fields are stripped before anything reaches the model or a cache.
3.0
Built around the existing business
3.1
Measured performance
Our October 2026 measurements recorded a median processing time of 2.7 seconds for account questions using the live model. Separate tests handled bursts of 100 simultaneous questions without failures, with the model stubbed.
3.2
Tested and refined with the client
Repeated test rounds with client staff shaped the product. QPOI investigated findings, delivered fixes and supported retesting, backed by more than 2,400 automated tests and separate evaluations against the real models.
3.3
Integrated with existing systems
The assistant uses the client's existing analytics and user accounts, with no account migration or separate credential store. QPOI deployed it on the client's infrastructure and provided a runbook for deployment and rollback.
4.0
Outcome
- Deployed on the client's infrastructure and in testing with the client's staff
- A new voice and chat product connected to existing analytics, user accounts and infrastructure
- Repeated client test rounds, with findings investigated, fixed and retested
- Backed by more than 2,400 automated tests and live-model evaluations
Behind the build / Engineering notes
How it works, in detail
How data access and answer checks work, the performance measurement methods, and the faults found and fixed during evaluation.
Built with
- TypeScript
- React (Vite)
- NestJS
- OpenAI Realtime API
- SQL Server
- WebRTC
- Docker
- nginx
Next project
Auddr.ai