The Industry's First
Network-Native
Inference Runtime.
Divide the payload. Aid the edge.
Not a cloud AI service. A distributed inference runtime. We shard open-weights SLMs across existing Cisco Catalyst 9300 Switches and CW917X Access Points via native Docker micro-containers. Leveraging asymmetric pipeline parallelism to run sovereign enterprise inference with zero dedicated GPUs.
Absolute Data Sovereignty.
Consider a financial institution processing highly sensitive PII. Employees using public, external AI tools create massive Shadow AI compliance breaches. Divaid keeps the prompt entirely localized, giving teams the AI they want while keeping the inference swarm physically sandboxed within the branch.
The 1,000-Node Compute Matrix.
Deploying multi-million dollar GPU racks in every branch or campus is financially and energetically impossible. But scale changes everything. By harvesting dormant background cycles from the Catalyst Switches and Wi-Fi 7 Access Points already hanging from your ceilings, divaid.AI turns your existing network into an on-premise AI cluster with zero net-new power draw.
The Hidden Supercomputer: On-Premise vs. Cloud GPU
Aggregating just 100MB per AP yields over 100GB of working memory—surpassing the 80GB VRAM of a dedicated $15,000 NVIDIA A100 GPU.
Instead of renting expensive AWS instances, DDA awakens the massive, dormant multicore compute power already hanging from your ceilings.
APs run 24/7 via low-power PoE. We scavenge stranded background cycles, bypassing the massive multi-kilowatt (kW) footprint of cloud server racks.
Zero cloud API calls. By providing a secure, internal LLM, we effectively cure the enterprise Shadow AI problem while ensuring corporate IP never leaves the firewall.
Real-World Flow: Asymmetric Pipeline Parallelism.
Processing a mission-critical 500-page Financial Audit or Live Network Triage without dedicated GPUs or cloud API egress. Rather than training a proprietary toy model, divaid.AI functions as an Asymmetric SLM Runtime & Partitioning Manager that shards standard open-weights models (Qwen-2.5, Llama-3.2, SmolLM2) across existing Cisco networking infrastructure:
1. Ingress Embedding (C9300)
Prompt tokenized in RAM; Ingress Embedding (W_embed) projects input tokens into initial activation vector x_0 (3.58 KB in FP16).
2. Sequential AP Pipeline Ring
Activation vector travels peer-to-peer across AP-01 → AP-02 → ... → AP-n over an isolated VLAN. Each AP executes 2 layers in 18 ms using bare-metal ARM NEON SIMD.
3. Zero-Disk-Write & Local KV-Cache
0 bytes written to flash (100% DRAM ring buffers). Localized KV-Cache reduces memory footprint by 12x (<61 MB RSS vs 96 MB cgroup ceiling).
4. LM-Head Projection & Sampling (C9300)
Final hidden state x_24 returns to C9300. High-speed AVX2/AVX-512 SIMD kernel projects d_model → V = 151,936 and samples the next token.
5. Multi-Swarm Task Recomposition
For massive audits (500 pages), multiple AP swarms (Vault, Teller, Risk) execute map-reduce chunks in parallel and merge into a sovereign report.
