Rabiyya Tahir

I am a final-year artificial intelligence student at Information Technology University, specializing in AI safety, LLM integration engineering, and model risk evaluation. Over the last two years, I have worked as an LLM Trainer and QA Developer at Turing (via Swift Solutions), contributing directly to high-stakes reinforcement learning pipelines (RLHF) and alignment datasets for leading research labs, including Anthropic and Meta.

My core technical achievements and capabilities include:

• Model Red Teaming & Alignment: Audited complex multi-app prompt architectures against strict 33-rule compliance frameworks. Experienced in identifying systemic risks, executing multi-turn prompt-response cycles, and resolving massive production data reworks to optimize logical correctness and structural safety.

• Adversarial ML & Guardrails: Designed a dual-layer, CPU-adapted visual prompt injection guardrail for my Final Year Project, benchmarking computer vision methods like PatchCore to detect embedded text instruction hacks at a 0.970 AUROC. I have also built full-stack, containerized text prompt injection layers using FastAPI, Docker, and active LLM judges.

• Agentic Workflows & Infrastructure: Built fully local, air-gapped autonomous multi-turn loops using Ollama (Llama 3.2 3B) with custom JSON function calling, and engineered 198 custom bag-level features from raw tabular columns for competitive machine learning ensembles (placing 3rd in CodeRush 2026).

  My deep focus lies at the intersection of socio-technical risk mitigation, algorithmic fairness audits (utilizing SHAP explanations), and engineering robust, automated testing rubrics to catch model vulnerabilities before they reach production.