Portfolio · AI security research

About PromptSlayer

A personal research project exploring how adversarial testing for language models can be made systematic, reproducible and useful to the team that has to ship the fix.

Why I built it

Most LLM safety checks assert on refusal keywords, which is easy to pass and tells you nothing. I wanted a harness that judges compliance semantically and produces a resilience number you can track across model versions.

Hardest part

Rate limits. Running thousands of probes against a provider means the scheduler, not the attack logic, decides throughput, so the worker pool is rate-limit aware, resumable and jittered rather than simply parallel.

What I'd do next

Agent-level testing: tool-call injection, retrieval poisoning and cross-session memory attacks, where the vulnerable surface is the orchestration layer rather than the model itself.

About this site

This is the portfolio presentation of the project, not a commercial product. The platform connects to a real backend engine to execute live audits against LLM endpoints, while also offering a safe, simulated demo mode for quick exploration.

Skills demonstrated

AI SECURITY
LLM red teaming, adversarial corpus design, jailbreak taxonomy
EVALUATION
LLM-as-a-Judge rubrics, confidence calibration, verdict triage
BACKEND
FastAPI, asyncio background tasks, rate-limit scheduling, SQLite
REALTIME
WebSocket event protocol, resumable streams, backpressure
FRONTEND
Vite, React, TypeScript, Tailwind, shadcn/ui
SECURITY UX
Verdict states, severity hierarchy, safe-by-default demos

Walk through it

Start with the safe demo, then read the dashboard and findings to see how a run turns into remediation guidance.

TRY SAFE DEMO