Cybersecurity senior (minor in Game Design) at High Point University. Air Force ROTC cadet, commissioning into the U.S. Space Force in spring 2027.
I build agent systems that run unattended against real services, and I spend most of my time on what they get wrong quietly.
A database or search backend fails, and the code above it reports "nothing there" instead of an error. The caller carries on as if the data never existed. I find this in open-source agent frameworks and the AI infrastructure around them, and send the fix upstream with a reproduction and a test that fails on the old code.
13 merged so far. Three of them are fixes to Feast, each merged by the same maintainer within nine hours of being opened.
| Project | What was wrong |
|---|---|
| haystack #3997 | A Qdrant backend failure came back as an empty result |
| haystack #4006 | In pgvector, an empty filter deleted the whole table |
| semantic-router #706 | A failed Qdrant read looked like an empty index, and remote sync then deleted local routes |
| agno #10682 | A delete that failed reported the run as missing |
| feast #6911 | Couchbase writes that failed were counted as written |
| feast #6918 | A failed Trino monitoring query read as "no metrics" |
| feast #6920 | A failed Dask monitoring read was reported as "no data" |
| cognee #5320 | Neptune graph writes reported success when nothing was written (my fix, merged through the maintainer's PR with co-author credit) |
| langchain-azure #1119 | SQL Server lookups that failed said "no such documents", and a failed delete could report success |
| langgraph-redis #206 | History "before" a checkpoint returned the whole history |
| langroid #1201 | A LanceDB write that failed was logged and the documents reported as ingested |
| redis-vl-python #752 | messages stopped at 10 or 1,000 messages instead of returning the whole session |
| mcp-server-manager #109 | Exported configuration was written with open file permissions |
I also have fixes open across many other projects, from LlamaIndex and Semantic Kernel to crewAI, Dify and LangChain. See them all.
If your project wraps a database, vector store, cache or queue, I will scan its adapters for this defect for free and send you a count and one finding worked in full. Fixes are a fixed price agreed before I start, and every one ships with a test that fails on your current code. eric.delorefice@outlook.com
unattended-agent-security - ten failures from running an LLM agent unattended against live systems, and the controls that prevent them. Nine of the ten reported success while failing.
A check that has never gone red proves nothing. Before I trust a fix, I make its test fail against the old code and read why it failed.
- Microsoft Certified: Azure Fundamentals (passed 30 September 2026)
- Interested in cyberspace operations, agent reliability, and the gap between a system that reports success and one that succeeded