01 / MODEL ADAPTATION
We have worked with LoRA fine-tuning.
Our engineering experience includes carrying out LoRA-based model fine-tuning and evaluating model-level adaptation. Fine-tuning experience should not be confused with deployment of a fine-tuned model in every production application.
Following LoRA fine-tuning work, the production Dayu assistant experience was implemented with a server-side prompt approach instead of serving the fine-tuned LoRA model. This describes Dayu's deployment choice, not abandonment of fine-tuning as a capability.
02 / SEPARATION OF RESPONSIBILITIES
Models generate language. Software enforces rules.
Prompts shape conversational behavior, but they are not a substitute for permissions, validation, data boundaries or deterministic application logic. Where a requirement must always hold, we design an application-level control rather than depending exclusively on an AI response.
03 / QUALITY ASSURANCE
Validate outputs against explicit checks.
Our GEO workflow engineering has included programmatic quality gates and regeneration paths for content that does not meet defined checks. The exact checks and human-review steps depend on the project; a model rating its own output is not treated as sufficient proof of correctness.
04 / KNOWLEDGE & ACCURACY
Separate source material from marketing claims.
We distinguish documented customer work from internal research, synthetic demonstrations and presentation artwork. A generated case study should not be published as an actual client delivery without supporting evidence. Assistants intended to answer from documentation should have a defined source scope and a path to human help when information is uncertain.
05 / HUMAN OVERSIGHT
Keep consequential decisions accountable.
Our current studio positioning does not offer AI as an autonomous decision-maker for consequential judgments about individuals. Potential high-impact uses require separate requirements, risk review and meaningful human responsibility, rather than being assumed safe by default.
06 / TRANSPARENCY
Label illustrative work honestly.
AI-generated artwork and simulated interfaces are identified as concept or demonstration material when they are not verified product screenshots. We do not present an unverified performance metric, customer review or case study result as fact.
07 / DATA & PRIVACY
Limit data exposure by design.
The default preview inquiry experience opens a local email draft or copies text rather than submitting it to a site database. A separately configured production API may process inquiry data with security and email providers. The privacy notice describes website-level third parties, including the Clutch badge, rather than implying that every visitor has zero external requests.
This is an engineering practice statement for the current Anvilstead site draft, not a certification, regulatory assessment or guarantee of identical behavior across all products. Before production publication, the company should confirm each operational claim against the deployed systems.