¶ AI Products — TeleSense & LLM Integrations
¶ TeleSense v1.0
AI-powered telecom analytics — on-premise, no data leaves your network.
TeleSense is Peon Technologies' flagship AI product, launching Q2 2026. It provides real-time and predictive analytics across the Namibia backbone using large language models for anomaly detection, capacity forecasting, and automated NOC summaries.
¶ Key Capabilities
| Feature | Description |
|---|---|
| Anomaly Detection | Real-time analysis of SNMP/NetFlow — flags unusual traffic patterns before they become incidents |
| Capacity Forecasting | ML-based utilisation prediction — 7/30/90 day forecasts per circuit |
| Automated NOC Summaries | LLM-generated shift handover reports from ticketing + monitoring data |
| VoIP Quality Analysis | MOS trend analysis from Homer 7 data — predicts degradation before customers notice |
| Client Health Scores | Per-tenant uptime, latency, and call quality scoring for QBR reports |
¶ Positioning
TeleSense v1.0 — On-Premise AI Analytics
┌─────────────────────────────────────────────────────┐
│ Data Sources │
│ ───────────────────────────────────────────────── │
│ Prometheus metrics ──► TeleSense Core Engine │
│ Homer 7 SIP data ──► (Python, FastAPI) │
│ NetFlow/SNMP ──► │ │
│ Vikunja tickets ──► ▼ │
│ LiteLLM Proxy │
│ (Claude Sonnet / Gemini) │
│ │ │
│ ▼ │
│ Insights Dashboard │
│ (React, internal) │
└─────────────────────────────────────────────────────┘
Deployment: On-premise, Kubernetes cluster (k8s-worker-01..03 at SWK DC1)
Privacy: All inference stays on-network. No telemetry to cloud providers.
¶ LLM Infrastructure
Peon uses a two-model strategy based on internal benchmarking (completed March 2026):
| Model | Provider | Use Case | Cost profile |
|---|---|---|---|
| Claude Sonnet 4 | Anthropic | Complex reasoning, incident analysis, report drafting | Medium volume |
| Gemini Flash | High-volume classification, tagging, summarisation | High volume, low cost |
¶ LiteLLM Proxy
All LLM API calls route through the self-hosted LiteLLM proxy at http://localhost:4000 (SSH tunnel for admin UI):
┌──────────────────┐ HTTP/OpenAI API
│ TeleSense ├──────────────────────►
│ OpenWebUI │ LiteLLM Proxy
│ Internal tools ├──────────────────────► (127.0.0.1:4000)
└──────────────────┘ │
┌─────────┴─────────┐
▼ ▼
Anthropic API Google AI API
(Claude Sonnet) (Gemini Flash)
Access LiteLLM admin UI:
ssh -L 4000:localhost:4000 root@102.130.71.211
# Then: http://localhost:4000/ui
¶ OpenWebUI
Internal AI chat interface at chat.peon.tech. Used by the engineering team for:
- Drafting incident reports and runbooks
- Code review assistance
- Network design Q&A
- Client communication drafts
¶ AI Benchmark Findings (Internal — March 2026)
Results from LLM provider evaluation across Peon-specific tasks.
| Task | Winner | Notes |
|---|---|---|
| Incident root cause analysis | Claude Sonnet | Best multi-step reasoning on log data |
| BGP config generation | Claude Sonnet | Accurate, fewer hallucinations on NANOG patterns |
| High-volume CDR classification | Gemini Flash | 10× cheaper, acceptable accuracy for bulk tasks |
| Customer support chat | Claude Sonnet | 82% first-contact resolution in pilot |
| VoIP quality summarisation | Gemini Flash | Fast turnaround, good enough for routine alerts |
¶ Upcoming: GPT-4o Customer Support Chatbot
Pilot underway — GPT-4o integrated into the customer support portal. First-contact resolution rate in pilot: 82%. Production launch target: March 20, 2026.
Architecture: Customer portal → GPT-4o (via LiteLLM) → Vikunja ticket creation on escalation.
¶ RAG Pipeline — Internal Knowledge Base
In-progress: Retrieval-Augmented Generation pipeline over internal documentation (NetBox exports, Vikunja history, network diagrams). Goal: NOC engineers ask natural language questions about the network and get accurate, cited answers.
- Embeddings: Stored in pgvector (PostgreSQL extension)
- Chunking: Per-page for Wiki.js content, per-device for NetBox
- Retrieval: Top-5 semantic search + keyword hybrid
- Generation: Claude Sonnet with citations