redcell
Defensive agent-robustness harness mapped to the OWASP Top 10 for Agentic Applications — deterministic oracles, a 146-case probe corpus, and a sandboxed three-level target.
status: active · 132 tests
sam@latino:~$ whoami
AI engineer. Agent systems, evals, and self-hosted LLM infrastructure — Rust + Python.
6 open-source projects · 524 tests · Rust · Python · TypeScript
~/projects
6 dirs-
-
longhaul
Early-adopter MCP 2026-07-28 release-candidate server in Rust — stateless core, the Tasks extension, and statelessness proven by round-robining one client across two server instances.
status: active · 67 tests
-
patchbay
Single-binary OpenAI-compatible gateway where privacy routing is enforced in the type system — a private request cannot select an external backend, by construction.
status: active · 13 tests
-
millstone
Clean-room BM25 + tree-sitter repo-map retrieval crate, with a bench harness against tantivy and SQLite FTS5 — the "you probably don't need embeddings (yet)" thesis.
status: active · 35 tests
-
callcheck
Tool-calling and structured-output conformance matrix for vLLM-served open models, with an 11-label failure taxonomy and a mock server that proves the scorer itself is correct.
status: active · 165 tests
-
eval-gate
Regression-eval CI gate — embedding-free scorers, drift detection with a k-repeat noise floor, and a sticky PR comment via a GitHub Action. Dogfooded by every other repo here.
status: active · 112 tests
~/writing
rss
BM25 beat my vector database (sometimes)
A crossover framework for lexical versus vector retrieval on code — and the adversarial bench harness I built so my own argument can lose.
Every model fails tool calling differently
Tool calling is the load-bearing primitive of every agent stack, and open models break it in at least eleven distinguishable ways. Naming the failure modes changes how you build the layer above.
Red-teaming my own agents with the OWASP Agentic Top 10
Turning "resists prompt injection" into a regression number: a deterministic harness, 146 probes across five OWASP agentic categories, and a hardening sweep that went 73% → 3% → 0%.
~/contact
- latinosammy2@gmail.com
- github
- github.com/slatino-dev
- linkedin.com/in/samlatino
- hf
- huggingface.co/SamLatino