Pelican - devsquad

Handling Closed-Source, Dynamic DOMs Across 3 AI Platforms:
ChatGPT, Claude, and Gemini Web use completely different, rapidly mutating SPAs. Standard page reloads and form submits don't exist here. We intercepted history.pushState to detect chat switches in real time, filtered out streaming assistant noise and nested elements, and built a universal clipboard fallback in our side panel so DOM changes never break the user experience.

Preventing "Self-Referential" Memory Ingestion:
When injecting past memories into the chat composer, the extension risked re-capturing its own injected context as "new evidence," creating a hallucinatory feedback loop. We solved this with strict boundary markers ([Context Passport: Memory]...) and pre-ingestion regex stripping that sanitizes prompts before extraction.

Beating MongoDB Atlas Vector Search Indexing Lag:
Atlas Vector Search indexes vectors asynchronously in the background. If a user saved or updated a memory and asked a question in the very next turn, Atlas wouldn't return it yet. We engineered a hybrid pipeline: querying Atlas while simultaneously computing in-memory cosine similarity on recent writes (< 5 mins) directly from MongoDB to deliver zero-lag recall.

In-Flight Composer Race Conditions:
If a user clicked "Use Memory" and continued typing while the network request resolved, blindly updating the composer would overwrite their newly typed thoughts. We built draft snapshotting that checks for in-flight edits upon response, cleanly merging memories without clobbering input.

Defense-in-Depth Secret Screening:
To guarantee credentials never leak into vector embeddings or logs, we built a two-tier screening pipeline (client-side pre-flight regex + backend inspection) with heuristic entropy matching that scrubs API keys, private tokens, and passwords before network transmission.