25-Year IT Veteran Builds LLM Middleware That Gates Memory, Filters Output, and Logs Everything
Neither-Beginning395 · reddit · 2026-09-16
A developer with 25 years in enterprise IT built a middleware layer that decides what an LLM may recall, checks outputs before delivery, and writes signed, hash-linked audit records of every decision.
Key points:
- Sensitive memories require re-authentication before recall
- The output filter has blocked the model leaking its own prompt in production, with the block itself logged
- Backed by OpenAI, Google and Anthropic models under zero-data-retention agreements
- First app is a bilingual English/Spanish wellness app called DeBrah; the layer itself is the point
- Live at miangel.ai, with the author explicitly inviting red-team attacks on the memory gate and the self-signed log's trustworthiness
More from coding & agent
- Tracking a FashionMNIST CV Project End-to-End with Weights & Biases — wandb · 2026-09-16
- BetterGravity: open-source mod adds in-app browser and BYOK to Google Antigravity — Yashjit · 2026-09-16
- prime agent v0.9.5 fixes bugs and picks its favorite update for its own release video — samsja19 · 2026-09-16
- Prompt history is the new source code: Reddit proposal for PromptSourceCode — niosurfer · 2026-09-16
- Why a domain knowledge graph beats prompts and horizontal agent platforms in agriculture — DevDminGod · 2026-09-16
- Early GLM video-input experiments: screen recordings for UI bugs, voice still weak — DevDminGod · 2026-09-16