production systems

The infrastructure
that runs the research.

The operating substrate behind the research: agent orchestration, corpus infrastructure, and memory architecture. Three production systems — each one designed, built, and operated to generate the data behind the research program.

Scope

Three production system architectures. Each one generates real data for the research program — not diagrams of hypotheticals.

System 01

Autonomous agent orchestration

Persistent workers routed across model families with shared state, memory, and real-world interfaces.

ORCHESTRATION ENGINE Job Queue · Task Router · Session Manager · Scheduler PERSISTENT WORKER POOL Claude Workers Opus · Sonnet · Haiku Code Workers Code Generation · Review Gemini Workers Flash · Pro · Image Gen Multi-Provider Pool 30+ Models On Demand MEMORY SYSTEM LanceDB Vector Store Core Memory · Archival · Procedures STATE & DATA SQLite State Store Projects · Tasks · Jobs · Logs SKILL MODULES Research · Analysis · Publishing Integration · Automation · Security EXTERNAL INTERFACES Messaging User Interface Email Triage · Route · Send Monitoring Observability Layer Browser Agent Web Automation HuggingFace External Data APIs Content Platforms Infrastructure APIs Device APIs
System 02

Media corpus pipeline

Built for Penelope Lawrence’s linguistics research. Full ingestion and retrieval path: capture, normalization, embeddings, vector indexing, and cited analysis across a 143K-headline corpus.

MEDIA INGESTION Outlets APIs Web Scrape Archives NORMALIZE Cleaning + Parsing Metadata Extraction EMBEDDING Headline + Context 1536-dim vectors CORPUS INDEX Embeddings + Metadata Hybrid: Dense + Sparse Outlet / Time Window ANALYSIS PATH RESEARCH QUERY Natural Language QUERY EXPAND HyDE · Multi-Query RETRIEVAL Top-K + MMR RERANKER Cross-Encoder Scoring LLM Context Window Packing Citation Grounding Hallucination Guards CITED RESPONSE Eval: Faithfulness · Coverage · Latency
System 03

Agent memory architecture

Core memory, archival facts, and procedures each do different jobs; the system works because those layers stay separate.

TASK Prompt / Event CORE MEMORY Identity · Constraints Budgeted Context PROCEDURES Matched skills · learned routines ACTIVE CONTEXT Open task · recent state ARCHIVAL MEMORY Facts · prior runs · retrieved recall Retrieve Summarize Consolidate CHECKPOINT Promote / discard MEMORY INDEX LanceDB · Task state · Fact store Each run updates facts, then promotes stable patterns into procedures.
Proof note

Live. These aren’t demos.

The research page describes the questions. This page shows the production systems that make those questions worth asking.

Research program