Hi, I'm Richard Wollyce

An open-source memory layer for fleets of AI agents, written in Rust and released under Apache 2.0. There is no embedding model anywhere in the retrieval path. It is a keyword index, a full-text index, and the code that fuses them. No network, and the same ranking on the same question every time. It can also tell you that nobody covers a question, instead of handing back the least wrong file.

• Retrieval runs two scorers rather than one. A keyword index built from the keys each file declares, and SQLite full-text search over the prose, fused with Reciprocal Rank Fusion. Then I measured which scorer wins which kind of question, because they are not interchangeable.
• Refusal is a first-class verdict here. On a question set authored blind and then checked adversarially, the deterministic layer alone declines to confidently answer 28 of the 30 out-of-scope questions.
• Warm route latency measures 0.68 ms p50 and 1.16 ms p95, in process. On a full 500-question run of LongMemEval-S it scores 97 percent on abstention, which that benchmark's own paper names as the ability memory systems fail hardest.
• Ships as an MCP server with four read-only tools, so Claude and anything else that speaks MCP reads the same base I do. There is deliberately no write tool a model can reach.
• Git is the privacy model, and there is no configuration file for it. A file git does not track is a file the system will not serve, and it refuses to open a base at all when git cannot be consulted.
• Roughly 17,000 lines of Rust across three crates, a single runtime dependency, over 200 tests, 33 architecture decision records. Every benchmark result carries the command, commit, machine and date that produced it.

The house I lead engineering for. It ships digital products and infoproducts, and my half is everything underneath the sale, from Pix checkout and attribution through server-side conversion tracking to the financial reconciliation at the end of the month. BiblinhaPlay is the oldest product still running in it, a cross-platform learning and entertainment subscription with roughly 500 users.

• I built the revenue instrumentation the house runs on. First-party browser and server conversion tracking, event deduplication, UTM attribution, revenue reconciliation, and the dashboards operators read before deciding what to run again.
• A generative-image pipeline with idempotent jobs, bounded concurrency, caching, persistence and telemetry. All of that exists so a failed job can retry without duplicating work or charging a customer twice.
• Product funnels deploy independently but share domain and UI packages, which is how payments, session recovery, attribution and backend integrations ended up living in one place.
• BiblinhaPlay I both ship and operate, for roughly 500 users, across a production web/PWA and an Expo/React Native client. Video streaming, music, printables, interactive games and gamification, all of it behind entitlement-based access.
• Billing runs on hosted subscription checkout with verified idempotent webhooks, plan-based entitlements and session-bound delivery of protected media. Content goes out through an immutable release pipeline that activates atomically and rolls back.
• BiblinhaCraft is a Three.js voxel world with deterministic terrain, progressive region streaming, versioned saves and touch-first controls.
• Digital checkout reaches physical production through address validation, print-ready processing, an operational queue and order-status workflows.

A cybersecurity learning path and resource hub for beginners, developers, and non-technical professionals.

• I wrote and organized the material myself.
• Modules cover fundamentals, threat analysis, password security and MFA, safe browsing, device hygiene, incident response, network fundamentals, and digital evidence handling.
• It runs on Mintlify, with a searchable documentation index and a path you can follow in order.