Defensive Technical Disclosures

A. On-device personalization of a speech-recognition model from user transcript corrections

A mobile application runs an on-device speech-recognition model for live transcription. When the user corrects a transcribed word, the correction is fed back to personalize the on-device recognition model for that device only — improving subsequent transcriptions — using the platform's on-device language-model-customization facility (supplying corrected text/phrases as boosted vocabulary), without synthesizing speech and without server-side training of that model. The result is a per-device, privacy-preserving transcription-accuracy loop driven by ordinary user edits. This disclosure covers the on-device transcription-recognition personalization only; cross-recording AI features are processed server-side and are not part of this disclosure.

B. Content-signature caching of a cross-recording AI briefing

A system generates a natural-language briefing (for example, an executive summary across all recordings in a project). It computes a content signature (for example, a hash) over the set of inputs contributing to the briefing. The generated briefing is cached keyed by that signature. On a subsequent request, if the current inputs produce the same signature, the cached briefing is served without re-generation; if the signature differs (a recording was added or edited), the briefing is regenerated and re-cached. This yields a briefing that stays current on its own while avoiding redundant generation cost, keyed to the actual content state rather than to a fixed time-to-live.

C. JSON-mode tool proposals with a confirmation round-trip (without native function-calling)

A conversational assistant proposes tool invocations by emitting a structured JSON object (via a language model's JSON / structured-output mode) rather than using a provider's native function-calling or tool-use API. Read-only tools auto-execute inline; mutating tools require an explicit user-confirmation round-trip before execution. The tool manifest is a fixed, frozen set; operator-only tools are filtered out of a non-operator user's tool context. This achieves tool-using assistant behavior with tighter control and provider-agnostic compatibility, with a safety gate on mutating actions.

D. Longitudinal per-speaker topic-trend tracking with percentage-change deltas

For each recurring speaker across a personal recording library, the system tracks topic and theme frequency over time and surfaces trend lines with percentage-change deltas (for example, "systems-design mentions up 47% this quarter"), together with per-speaker stylistic signatures (for example, a hedging rate). The computation is performed over the user's own content.

---

These disclosures are published by Bonfiyah, Inc. solely to establish prior art and prevent third-party patenting of the described subject matter. Publication does not grant any license to Bonfiyah's trademarks, source code, or other intellectual property, and is provided "as is" without warranty of any kind.