You cannot ship an LLM proxy without evals
Thesis note: baseline versus protected profile, 17/17 PASS, and what that number does not prove.
A replayable suite — layer hit, HTTP status, latency — instead of a screenshot of a chat.
OpenAI-compatible reverse proxy: inbound threat detection, reversible DLP, routing, and evals. MSc thesis deposited (pending defense).
Drop-in middleware between apps and any LLM provider. Pipeline: sanitisation and OWASP-style filters, LLM-as-judge in parallel with DistilBERT, bilingual DLP, then LiteLLM routing. Evaluation battery 17/17 PASS; ~175 pytest. Repo: github.com/r4fik1/ai-security-gateway.
Thesis note: baseline versus protected profile, 17/17 PASS, and what that number does not prove.
A replayable suite — layer hit, HTTP status, latency — instead of a screenshot of a chat.
Security as middleware: an OpenAI-compatible reverse proxy for LLM traffic.
If every client talks to the same Base URL, notebooks, chatbots and batch jobs share one pipeline.
High-assurance rekey / generate-root as a tested Go product, with GPG and YubiKey human gates.
Enterprise Vault operations that used to live in tribal shell scripts: one cross-platform binary, health and token checks, coverage gate ~95%, SSDLC in CI. Internal Swisscom tooling — no public repo.
From team-level metrics to a scheduled data product for leadership — without spreadsheet archaeology.
Python pipelines (Parquet, object storage, scheduled BI) covering training, assessments and vulnerability posture. Fail-closed publishes; cloud access via short-lived OIDC, not long-lived keys.