Custom AI Corpus · 3 weeks to production
Gerolamo | Building a Custom AI Corpus in Three Weeks
Adjective (Internal Proof System)
Adjective built Gerolamo as a public proof of the custom AI corpus model, demonstrating that a small team running a focused Field Discovery motion can ship a production system in weeks, not quarters. Started April 5, 2026. Public surface live within two weeks. The access logs from the first three weeks told a story we did not expect.
- 600 unique visitors with zero paid acquisition
- 2,152 MCP agent calls in 3 weeks
- 5,204 defensibility scans from 399 browsers
- 525 technical domains tracked continuously (up from ~190 at launch)
The challenge
The open-source technology ecosystem moves too quickly for manual research. Millions of new repositories, research papers, and AI models appear every month. Existing tools only reflect surface popularity, such as GitHub stars, download counts, and trending lists. They fail to address strategic questions regarding defensibility, impending commoditization by frontier labs, and genuine technical novelty. Gerolamo was built to evaluate durability, risk, and novelty simultaneously, making those metrics queryable by both analysts and AI agents.
The solution
Built Gerolamo, an agent-native intelligence corpus covering roughly 190 domains with scored sources, semantic search, relationship graphs, and a 31-tool MCP server
Approach
01
Corpus Architecture & Ingest Pipeline
Designed a multi-source ingestion system that normalizes code, research, and models into a single searchable landscape with daily refresh cycles
- GitHub, arXiv, and Hugging Face ingestion pipelines
- Entity normalization across ~190 technical domains
- Daily pipeline: ingest → evaluate → fuse → materialize → embed
- Junk filtering and deduplication layer
02
AI Evaluation & Scoring Engine
Built an LLM-powered evaluation layer that scores every entity on defensibility, frontier risk, novelty, and composability
- Defensibility scoring (1-10) with full reasoning chains
- Frontier risk assessment (LOW / MEDIUM / HIGH)
- Novelty classification (breakthrough → derivative spectrum)
- Sources: structured, typed, queryable units representing GitHub repos, arXiv papers, and Hugging Face models
03
Agent-Native Surface (MCP Server)
Exposed the entire corpus as a Model Context Protocol server so AI agents can search, score, compose, and synthesize intelligence alongside human analysts
- 31 MCP tools for agent-driven research workflows
- Semantic search (1024-dim OpenAI embeddings)
- Workspace composition with 5 synthesis modes
- Compatible with Claude Code, custom GPTs, and any MCP client
04
Public Surface & Relationship Graphs
Shipped the human-facing interface with ranked views, entity drill-downs, creator profiles, trajectory charts, and the Analytics macro-intelligence layer
- Defensibility and trending ranked views
- Entity detail pages with relationship graphs
- Creator authority scoring and portfolio views
- Analytics: AI-generated topic-level summaries across all tracked domains
Impact
- Time to Production
- 3 weeks
- From zero to a live, scored, searchable intelligence corpus covering ~190 technical domains
- Agent Adoption
- 2,152 MCP calls
- The MCP server became the second busiest endpoint on the platform, ahead of every human-facing workflow. Agent traffic outpacing human traffic per-IP and widening week over week.
- Organic Reach
- 600 unique IPs
- Zero paid acquisition, no launch plan, no ads. Posted a few links and went back to building. MCP surface used by agents inside organizations Adjective did not introduce itself to.
- Defensibility Scans
- 5,204 visits
- 399 unique browsers averaging 13 page loads per visitor on the defensibility rankings alone, indicating strong return-visit behavior in week three
“Humans browse. Agents query. The ratio between the two seems to be widening week over week, and the agent surface is starting to feel like a first-class product on its own.”
Next case study
Snap | Visual Annotation for Agentic Coding
Your workflow could be next.
Book 30-Minute Briefing