AURAS INNOVATION · PRIVOID.AI
AI and Results Disclosures
Effective 30 September 2026 · Version 2
Privoid.ai shows reference builds to explain possible enterprise AI approaches. The figures on this site are not verified customer outcomes, independent certifications, contractual service levels, or promises that another workload will produce the same result.
Read a large number with its basis. “Measured” means observed in a stated test setup. “Modeled” means calculated using assumptions or third-party price cards. A percentage saving in running cost is not a total-cost saving, a quote, or a guarantee.
1. Status of the examples
The site describes these as internal reference builds or prototypes, not client deployments or endorsements. The following summary reflects the status shown on each case-study card and the supporting descriptions; it does not replace the underlying test report.
| Example | Stated basis | Important limit |
|---|---|---|
| EdgeVision Inspector | Measured on one RTX 3060; cloud comparison modeled. | A reference line and test conditions do not establish performance or safety on a customer's production conveyor. |
| FlareGuard AI | Reference prototype using simulated cameras; cost modeled. | Not evidence of deployment or approval at an oil and gas facility. |
| Ghost | Measured on an RTX 3060 12 GB; cloud cost modeled. | “$0 API bill” does not mean zero hardware, power, support, or labour cost. |
| Hivemind Review | Some figures measured and some modeled, as labeled in the case. | Review quality, code privacy, and cost depend on the repository and deployment architecture. |
| Support-Ticket Triage | 100 hand-authored tickets; cost modeled. | That sample does not prove equal quality on real customer tickets, other domains, or at the modeled volume. |
| DocExtract-Zero | 26 held-out invoices; cloud cost modeled. | Small held-out samples and public-dataset layouts do not establish production accuracy on a client's documents. |
| RiskLens | Measured on one consumer GPU; cost modeled. | Not an insurance underwriting approval or an automated decision system. |
| Covenant | Measured on a sample contract; cost modeled. | Not a legal opinion or evidence of accuracy across contract types or jurisdictions. |
2. Cost and savings methodology
Published API prices can change or differ from a customer's contracted rate. Modeled comparisons may use estimated token volumes, throughput, hardware amortisation, energy, and support. Unless the individual case explicitly includes them, one-off engineering, migration, data preparation, evaluation, security review, licences, taxes, downtime, and quality rework are outside a running-cost comparison. Different volume or model-quality requirements can reverse a saving. An estimator output is an illustration, not a quotation, financial forecast, or procurement recommendation.
3. Performance and quality
Latency, accuracy, and reliability results apply to the stated sample, environment, hardware, software version, and measurement method. A median is not a worst-case result; a score on authored or public test data is not the same as performance on customer data. Benchmark labels and ground truth may contain errors. Before any client commitment, the parties should agree representative data, baselines, pass/fail thresholds, quality review, and acceptance tests in writing.
4. Privacy, security, ownership, and regulated work
“Local”, “on-premises”, “air-gapped”, or “zero egress” describe possible architectures or a particular reference test; they do not guarantee that every deployment has the same boundary. Hybrid routing can send selected data to an external provider if expressly authorised in the signed design. A network check is not an independent security certification. Ownership of code, models, weights, data, and third-party components is determined by the signed agreement and applicable licences. See Security and Data Handling.
Reference builds are not approved for autonomous medical, legal, financial, insurance, employment, public-service, industrial-control, or other high-impact decisions. They require domain-specific validation, human oversight, and applicable approvals before any production use. See Responsible AI and Acceptable Use.
5. Commercial statements
A free teardown, founding-partner pilot, benchmark target, milestone, or no-build-fee condition is only an invitation to discuss a possible engagement. Eligibility, scope, baseline, exclusions, fees, remedies, and data handling exist only when written into a signed proposal or contract. We will not present a reference prototype as a paid customer result. For the methodology behind a particular figure, email support@privoid.ai.