Complex AI systems, not just simple chatbots

Ciphr Labs builds AI that can listen, speak, search knowledge, trigger workflows, book calls, and hand work to humans. The assistant on this site is the first working layer: it can chat, use voice input, speak back, collect appointment requests, and connect to a local Qwen model through Ollama.

What we build

AI systems that sit inside real business workflows

A useful assistant is more than a reply box. It needs memory, company data, voice, permissions, handoffs, logging, and a way to notify the team when a visitor is ready.

Voice assistants that qualify leads, answer questions, and book appointments.

RAG systems that search company documents, policies, tickets, contracts, and knowledge bases.

Workflow agents that trigger tools, update CRMs, send summaries, and hand off to humans.

AI dashboards that monitor operations, detect patterns, and surface decisions.

Private model deployments using open source models where ownership and cost control matter.

Talks on the site

Visitors can open the bottom right assistant, type questions, speak through the browser, and hear replies.

Books appointments

The assistant captures name, email, preferred time, and project goal, then saves the lead for follow up.

Connects to tools

The next step is webhooks, CRM updates, email alerts, calendars, Slack, and internal dashboards.

Open source stack

No Vapi lock in

The production path uses free and open source tools where possible. Vapi is not required.

Dograh

Closest open source Vapi style platform. It is self hostable, designed for production voice agents, and supports visual workflows instead of locking the project into a paid voice wrapper.

View source

LiveKit Agents

Best infrastructure layer for browser voice because it handles realtime audio, WebRTC rooms, interruptions, and server side voice agents.

View source

Pipecat

Best pipeline layer when we want deep control over speech to text, LLM routing, text to speech, tool calls, and specialist handoffs.

View source

Qwen through Ollama

Good open model path for reasoning, code, and company knowledge. The site widget can connect to a local Qwen model through Ollama.

View source

Integration plan

From site widget to human style calls

The site widget is live now. Production voice calls need the same brain connected to realtime audio, telephony, tool calls, and notifications.

Bottom right site assistant that answers visitors, explains services, and collects project details.

Voice mode in the browser for speaking to the assistant and hearing human style replies.

Booking intake that stores appointment requests and project leads for the Ciphr Labs team.

Production voice calls through Dograh or LiveKit with Pipecat pipelines and Qwen model routing.

Recommended production stack

Use Dograh when we want the closest self hosted Vapi style platform. Use LiveKit Agents and Pipecat when we want custom realtime voice infrastructure. Use Qwen through Ollama or an OpenAI compatible server for the open model layer, then add webhooks so booking requests reach the team automatically.

Current site widget works with `/api/assistant`, `/api/lead`, browser speech input, browser speech output, and optional `LEAD_WEBHOOK_URL` alerts.

Build the assistant