EdgeVoC

Problem

Sending customer feedback to a cloud-hosted AI API creates privacy exposure — data leaves the organization and is processed by a third party. For organizations with strict data residency, privacy, or cost constraints, this limits the use of AI-assisted analysis.

The question: can structured AI analysis of customer feedback run entirely on-device, without any call to an external model during inference?

Product concept

A local edge-AI workflow that processes customer feedback entirely on-device during inference. The system takes raw feedback records, runs a local model managed by the QVAC SDK, and produces structured output covering themes, pain points, risk signals, recommendations, and an executive brief.

No external AI API is called during inference execution. The analysis stays on the machine where the tool runs.

How the system works

• Takes customer feedback records as input (15 records in the local execution evidence run)
• Invokes a locally managed model via the @qvac/sdk runtime (LLAMA_3_2_1B_INST_Q4_0)
• Performs inference entirely on-device — no external API call is made during this step
• Produces a structured analysis report covering: Themes, Pain Points, Risk Signals, Recommendations, Executive Brief
• Executed on an HP ProBook 455 G9 running Windows 11 (CPU-only, no GPU required)

Main conceptual workflow

The diagram below is a conceptual illustration of the system flow. It does not reproduce internal documentation.

My contribution

• Designed the product concept and the privacy framing for local edge-AI analysis
• Selected and configured the local model and QVAC SDK runtime
• Structured the input/output flow for customer feedback analysis
• Validated output quality and report structure against the 15-record test set
• Prepared submission materials; submitted at frozen commit b26663b

Technical implementation

• @qvac/sdk — local model runtime
• LLAMA_3_2_1B_INST_Q4_0 — locally executed, quantized language model
• HP ProBook 455 G9 / Windows 11 — execution hardware (CPU-only)

Evidence and validation

• Local execution evidence exists: the workflow ran entirely on-device during evidence capture
• All 15 feedback records were analyzed without an external API call during inference
• A recorded demonstration exists (URL not available in verified source materials — not reproduced here)
• Public repository available at frozen commit b26663b for code inspection
• Submission is frozen: no changes have been made to the repository since the submission commit

Limitations and transparency

• Local proof of concept: this is not a production customer-data platform
• Small test set: the 15-record evidence run reflects a validation scenario, not a large-scale deployment
• Model capabilities: LLAMA_3_2_1B_INST_Q4_0 is a compact quantized model; output quality differs from large cloud-hosted models
• No production privacy certification: this page does not claim production-grade privacy compliance or data residency certification
• Submitted / frozen: no further development is planned following submission