The archive drawer.
Deep-dive system notes — architecture, decisions, tradeoffs, and what actually broke. Written for people who like the machinery.
Routing Before Reasoning
Pre-routing keeps the full tool catalog out of the expensive reasoning model's context. In a 100-tool benchmark, top-5 routing held downstream context nearly flat and sharply reduced priced inference cost, while exposing a separate tradeoff between routing coverage and final tool selection.
SmartInvoiceExtractor
Regex extracts high-confidence fields from pdfplumber-parsed PDFs, while Gemini handles missing or ambiguous values. A quality gate, normalization layer, and regex suggestion loop help control when model inference is used and identify extraction patterns for further review.
Memora
Memora explores how AI-agent execution can be captured, inspected, and independently verified. Its public components combine local activity capture, signed evidence records, shared verification primitives, and a developer SDK for working with execution and memory data.
Retrieval Adaptation Pipeline
The linked final report describes a 6,044-example YouCook2-derived procedural disruption benchmark and evaluates a 390-example visual missing-ingredient test subset. A config-driven pipeline compares text-only and vision-conditioned generation with retrieval and reranking, while persisting generation and evaluation traces.
TrustGraph
Four-predicate, regex-based screening for Solidity trust-boundary risks, with optional Foundry PoC tests, Markdown/JSON reports, and VS Code diagnostics. Deterministic rules set the findings and severity; optional model-generated explanations follow the scan.
AgentOS
Git worktrees, command policy checks, diff inspection, and a Rust-backed desktop runtime support supervised coding-agent workflows. A SQLite store persists graph, event, and session records; the browser preview simulates native operations and uses an in-memory store.
Sageo
A verifiable identity and interaction layer for A2A-protocol agents — Coco smart contracts on MOI register agent identities and store hashed interaction proofs; a TypeScript SDK wraps A2A clients and handlers to inject trace metadata without storing raw payloads.
Clinical Speech ML Pipeline
End-to-end speech analysis pipeline: Expo mobile capture, FastAPI ingestion, FFmpeg normalisation, Whisper transcription, Librosa acoustic metrics, and Postgres persistence with an analytics endpoint. Prototype-stage — no clinical validation.