Naledi Sithole
en-ZACalmPatient
0:07
Naledi handles long-form support calls with a measured, patient delivery.
All products
Agent platform
The platform your AI agents run on.
The NPU Labs Platform is where agents are built, deployed, and supervised. Agents and the workflows behind them, the MCP tools and knowledge bases they reach, the voices they speak with, and the guardrails they answer to, all in one console. A full contact center is one of the workloads it runs, across WhatsApp, voice, web chat, email, and internal channels. Self-hosted on your NPU Labs node, or hosted by us.
Channels covered
One platform
WhatsApp
Forms, voice notes, documents, status updates.
Voice
Inbound and outbound calls over SIP or WebRTC.
Web chat
Embedded chat on customer-facing properties.
Email
Triage, structured replies, case routing.
Internal
Agent-assist for human teams already on the floor.
What the platform handles
Customisable WhatsApp formsMulti-channel conversationWorkflow automationBusiness system integrationIdentity verificationKnowledge base + RAGHuman handoverGuardrails + safetyObservabilityHosted, self-hosted, or hybrid
The stack the platform is assembled from
WhatsAppGmail
NVIDIA NeMo
LangGraph
Qdrant
MCP
Deepgram
ElevenLabs
Ollama
PostgreSQL
Grafana
DockerInside the console
The platform your team runs it from.
The contact center is one workload on the NPU Labs platform. The same console carries the agents, the workflows behind them, the tools they reach, and the guardrails they answer to.
01
One console for the whole platform
Monitoring, integrations, the agent workbench, and the guardrail layer in a single shell, scoped per tenant.
02
Every agent, with its own numbers
Agents deployed, concurrent sessions, and session minutes for the billing period, with latency, throughput, and error rate per agent.
03
Workflows you can read step by step
Each workflow shows its flow, the escalation branch, and a step breakdown with p50, p95, throughput, retries, and error rate.
04
Tools and knowledge you control
MCP servers, HTTP tools, and knowledge bases are registered in the console, so what an agent can reach is configuration rather than code.
05
Guardrails as rules, with a trigger count
Input, output, intent, and policy rules each carry a scope, an action, and how often they fired, so governance is something you can point at.
How it works
Four layers, one audited path.
The workflow engine and the rules engine own the verification decision. Every turn flows through the same gated path, on the same node, in your environment.
Layer 1 · Channel
Voice
WhatsAppWeb chat
EmailLayer 2 · Orchestration
LangGraph orchestrator
Owns state, verification level, escalation paths
Layer 3 · Reasoning + retrieval + tools
Reasoning LLMQwen 3 / Nemotron
RAGQdrant + bge-m3MCP toolsCRM, rules, PostgresLayer 4 · Safety + audit
Answer & log
Escalate / human handover
Refuse & audit
NeMo Guardrails gates every turn. Full transcript, verification level, and tool calls land in the audit ledger.
Platform capabilities
Everything an agent needs, in one platform.
Voice, chat, identity, knowledge, integrations, safety, and observability are the same product. You configure each agent. The platform handles the rest.
Multi-channel
WhatsApp, voice, web chat, email, and internal channels handled from a single configuration. WhatsApp forms are a first-class feature, not an add-on.
Identity verification
WhoYou biometric, PIN, and knowledge-based verification, with verification level gating which tools and data the agent may access.
Knowledge base + RAG
Per-agent knowledge bases with bge-m3 embeddings and reranking. The agent answers from approved sources, not from training memory.
MCP tool access
Connect to your CRM, rules engine, document store, and PostgreSQL workflows through Model Context Protocol servers you control.
Guardrails + safety
NeMo Guardrails plus a Nemotron content-safety model. Refusals, scope limits, and PII rules are policy, not prompt suggestions.
Multilingual
Deepgram Nova-3 STT and ElevenLabs TTS handle English. Spitch provides af-ZA. Add a language by adding a model, not by rewriting the platform.
Human handover
Failed verification, vulnerable customers, regulated decisions, and unclear answers route to a human team with the full captured context.
Evaluation built-in
DeepEval scores conversations for quality, safety, and task completion. DeepTeam red-teams the agent against jailbreaks and PII leakage.
NPU Voice
Deploy AI agents in over 100 languages, on voices you can configure or clone.
Persona is a first-class configuration on every agent. Pick a voice that fits the brand, the channel, and the customer. Clone an internal voice when you need consistency across regions.
What it does
Real business processes, completed in the channel.
Not a chatbot that only replies. NPU Labs Platform runs the repetitive, structured conversations your support, sales, and back-office teams handle by hand today.
Customer support
Answer common questions, classify the request, collect details, and open a case, or hand over with context.
Sales & lead qualification
Capture leads, ask qualifying questions, collect requirements, and route to the right sales process.
Onboarding
Guide customers through onboarding steps, collect information and documents, and submit to internal systems.
Claims intake
Capture claim details, request supporting documents, confirm the information, and create a structured case.
Document collection
Request and receive identity documents, proof of address, payslips, signed forms, and claim evidence.
Verification flows
Walk customers through PIN checks, identity questions, biometric links, or external verification services.
Service requests
Capture cancellations, upgrades, complaints, bookings, account changes, and policy updates.
Internal process capture
Let staff, partners, and field teams submit structured information through chat instead of a separate portal.
The admin shell
One screen to run the floor.
Configure every agent without touching code. Search any conversation, replay it against a different model, review every flagged turn. Operational insights for the people running the floor, alongside the engineers shipping it.
- Search conversations
- Replay and compare
- Supervision events
- Operational insights
npu-labs · consoleLive
Agents
- Inbound voiceLive
- Onboarding WhatsAppLive
- Claims intakeLive
- After hoursDraft
Inbound voice · Config
- Language modelQwen 3 32B
- Voice personaSA-English (custom)
- Speech-to-textDeepgram Nova-3
- Knowledge bases3 attached
- MCP tools5 connected
- GuardrailsNeMo + Nemotron-Safety
- Routing rules12 rules
- EvaluationDeepEval · weekly
Supervised by design
A supervisor agent watches the core agent.
NPU Labs Platform never runs unwatched. A separate supervisor agent monitors the core agent on every turn, scores each response against the standard AI quality and safety metrics, and corrects the core agent the moment something drifts. Issues are caught and behavior is fixed as the conversation happens, not surfaced in a report weeks later.
Quality
- Helpfulness: Does the reply actually help the customer?
- Relevance: On-topic and on-task for the request.
- Correctness: Factually accurate against the source of truth.
- Coherence: Clear, consistent, and easy to follow.
- Completeness: Nothing important left out.
- Instruction adherence: Stays within the brief and the rules it was given.
Safety
- Toxicity: Offensive, abusive, or harmful language.
- Bias: Unfair or skewed treatment.
- Hallucination: Claims not grounded in approved sources.
- PII leakage: Personal or sensitive data exposed.
- Prompt injection / jailbreak: Attempts to override the system.
- Off-policy drift: Straying outside the approved scope.
Experience
- Sentiment & tone: How the customer is reacting in the moment.
- Task completion: Did the process actually finish?
- Latency: How fast the agent responds.
Monitor behavior in real time.
When a behavioral score crosses the line (a toxic phrase, a hallucinated policy, a reply drifting off-scope), the supervisor steps in before the customer feels it, and logs the event for review and the next retraining cycle.
Open-source first
Built on the open-source stack you can audit.
No hidden inference. No black-box vendor. Every component has its own documentation, its own license, and runs on your node.
LiveKit Agents
Real-time voice transport
LangGraph
Conversation orchestration
Qwen 3
Reasoning + tool calling
Nemotron
Alternate reasoning + safety
Deepgram
Speech-to-text
ElevenLabs
Text-to-speech
Spitch
af-ZA STT and TTS
NeMo Guardrails
Policy and refusal
Qdrant
Vector store
bge-m3
Retrieval embeddings
DeepEval
Quality scoring
DeepTeam
Red-teaming
Grounded in research
Built on the work of others.
We are a lab. What we build is grounded in published research.
- [01] NeurIPS · 2020
Retrieval-Augmented Generation for Knowledge-Intensive NLP Tasks
Lewis et al. (2020)
Backs: RAG retrieval layer. The technique we ship verbatim for per-agent knowledge bases.
- [02] ACL Findings · 2024
M3-Embedding: Multi-Linguality, Multi-Functionality, Multi-Granularity Text Embeddings Through Self-Knowledge Distillation
Chen et al. (2024)
Backs: Embedding model. bge-m3 is named on the stack, this paper is its source.
- [03] IEEE TPAMI · 2018
Efficient and Robust Approximate Nearest Neighbor Search Using Hierarchical Navigable Small World Graphs
Malkov and Yashunin (2018)
Backs: HNSW index. The algorithm Qdrant uses to retrieve from the agent knowledge base.
- [04] ICLR · 2023
ReAct: Synergizing Reasoning and Acting in Language Models
Yao et al. (2023)
Backs: LangGraph orchestrator. Built on the ReAct interleaved reasoning and tool-call pattern.
- [05] EMNLP Demo · 2023
NeMo Guardrails: A Toolkit for Controllable and Safe LLM Applications with Programmable Rails
Rebedea et al. (2023)
Backs: Safety layer. We run NeMo Guardrails in production for refusals and scope limits.
- [06] NeurIPS Datasets and Benchmarks · 2023
Judging LLM-as-a-Judge with MT-Bench and Chatbot Arena
Zheng et al. (2023)
Backs: Supervisor agent. The LLM-as-judge formulation we ship for live quality and safety scoring.
- [07] EMNLP · 2023
G-Eval: NLG Evaluation Using GPT-4 with Better Human Alignment
Liu et al. (2023)
Backs: Supervisor metrics. DeepEval's G-Eval, which the supervisor uses to score each turn.
Citations include preprints alongside peer-reviewed work. We interpret current LLM-agent research alongside established cognitive science, databases, and information retrieval.