What we build,
and how we build it.

We were building internal AI agents that read codebases and answered questions about them. The agents worked, but the token bills didn't make sense. We traced it to one problem: context waste. Every time an agent needed to understand a module, it consumed entire files, directory listings, and search results — 40 to 70 percent of the tokens contributed nothing to the answer.

So we built mnemosyne-engine: a retrieval engine that indexes a codebase and returns only the chunks that matter, ranked by value-per-token, within a configurable budget. The engine runs a hybrid pipeline — BM25, TF-IDF, symbol matching, usage frequency, and four additional signals — all fused through Reciprocal Rank Fusion and re-ranked by a cost model that penalizes low-information content. A four-stage compression pipeline strips boilerplate while preserving function signatures, control flow, and documentation. We benchmarked it against a real 844-file production codebase: 74–78% fewer tokens, 92–98% faster queries, equivalent answer quality.

Then we built the hands. mnemosyne-mcp is a Model Context Protocol server that plugs the engine straight into Claude Code, Cursor, Zed, and any other MCP-compatible host. mnemosyne-ollama is a lightweight MCP host that runs the engine fully local against any tool-capable Ollama model — Llama, Qwen, Gemma, Mistral, Phi — with one command and zero configuration. Same engine, three delivery vectors, no cloud required in any of them.

All three are open source under AGPL-3.0 (commercial licenses available), pure Python, zero runtime dependencies beyond their direct protocol needs. 293 tests in the engine alone. Published with OIDC trusted-publisher attestation on PyPI. No telemetry, no analytics, no phone-home. The developers building the next generation of AI tools shouldn't have to solve context waste from scratch — and they shouldn't have to send their code to a third party to get the benefit.

Medicare Advantage organizations deal with enormous volumes of clinical documentation. Coding teams work through scanned PDFs — hundreds of pages per chart — looking for diagnosis codes that support risk adjustment. It's meticulous, slow, and the stakes are real. A missed code means lost revenue. An incorrect code means audit risk.

The existing workflow was almost entirely manual. Coders scrolled through poor-quality scans, mentally tracking conditions across dozens of encounter notes. There was no structured way to surface evidence for negated or historically mentioned conditions. And sending any of this data to a cloud service was off the table — protected health information stays on-prem, full stop.

We built an on-premise pipeline that ingests chart PDFs, applies text extraction with OCR fallback, assesses page quality, and runs clinical code detection with explicit negation modeling. Every detected code maps to the relevant risk adjustment category using the organization's selected payment year model. Evidence is extracted with page-level and character-offset provenance — so a reviewer can trace any finding back to exactly where it appears in the original document.

The key design choice was conservatism. When a binding is ambiguous, the system flags it rather than asserting it. Reviewers work from a structured report of candidates and flags, not a blank page. Their time shifts from searching to validating — and every finding is independently auditable.

Compliance teams at regulated organizations — healthcare operators, financial services firms, technology companies handling personal data — are responsible for tracking a fragmented landscape of federal guidance, state-level privacy laws, and enforcement actions. It's a lot. Sources are scattered, formats vary, and by the time a material change shows up in informal channels, the window for response may already be closing.

Manual monitoring works until it doesn't. Someone misses a Federal Register update. A state AG issues guidance that nobody catches for three weeks. An enforcement action establishes a new precedent that affects your deployment model. The cost of these misses isn't theoretical — it's operational.

We built a system that monitors over 50 federal and state regulatory sources daily. Changes are classified by urgency, filtered to the subscriber's configured industry and state profile, and delivered as structured weekly briefs. Each item includes a plain-language summary, a color-coded action checklist, and the original source reference. Urgent items surface immediately as priority alerts.

The result is that compliance and legal teams work from a structured action list rather than raw documents. High-urgency changes — new conformity requirements, final rules, enforcement guidance — surface immediately. Teams respond faster because the signal-to-noise ratio is dramatically better. And every item has full source attribution, so nothing is taken on faith.

Source attribution on every item. Urgency classification with explicit criteria. Operator-configurable industry and state filters. No assertion without plain-language explanation. Structured action checklist format that enforces human review before action.