AI Work
AI that shows its receipts.
Scholarship cannot tolerate a confident guess. An LLM that invents a citation is not a convenience problem — it is a forgery. So my AI work starts from a different premise: generation is the easy part; verification is the system.
Everything I build follows the same discipline — cryptographic proof binding claims to sources, deterministic pipelines where an agent reasons but code keeps the ledger, and a human approving every external action.
350,000+
words of patristic translation
80+
lectures processed through the AI pipeline
6
open-source tools in production
9
ancient languages in the toolchain
Featured System
CiteAgent
github/Areopaguaworkshop/citeagent ↗A research agent that answers only with what it can prove.
A hybrid Rust + Python system: research materials are indexed into a Merkle-verified knowledge base; every generated claim maps to a specific text passage verified by SHA-256 with a full Merkle proof. Retrieval is plain BM25 — auditable and cheap — while a Rust (ratatui) TUI hosts interactive sessions. Published on PyPI as cite-extractor.
Seven deterministic agents
Grounded Media Pipeline
From lecture hall to bilingual archive
The production line behind GCDFL's 80+ lectures: Whisper transcription, Wenbi converting recordings into structured Markdown, bilingual translation with DeepL first and LLM fallback, then vedit rendering subtitles and uploading to YouTube with automated metadata and playlists. Human-reviewed output only.
Agent-Native Systems
JobFinder
An agent-native job-search system built on an unusual split: deterministic Python owns the plumbing — CV ingestion, job-source normalization, SQLite storage, ATS and legitimacy signals, application tracking — while the agent supplies reasoning through portable SKILL.md skills shared across coding agents. No embedded LLM, no hidden calls. Outputs are drafts and checklists; the agent never submits or sends anything without human approval.
AI Stack