My Journey So Far

Trajectory Labs | AI Safety Workspace Canada

• Designed LLM evaluation pipelines using judge-router architectures to assess reasoning quality, sycophancy, and behavioral consistency across prompt variants and model configurations.
• Built and iterated on RAG systems with LangChain and LlamaIndex, experimenting with retrieval strategies and contextual grounding to measure downstream effects on model reliability and misgeneralization.
• Developed prompt-based behavioral testing frameworks probing alignment-relevant behaviors including agreement bias, instruction following, and reasoning robustness.
• Publishing research notes and technical breakdowns to contribute to open AI safety discourse and knowledge-sharing.

Futurekind AI

• Designing and building an AI Extinction Tracker web platform analyzing how AI-driven automation and agentic systems contribute to ecological and animal welfare risks.
• Developing a structured risk-mapping framework connecting AI deployment pathways to downstream environmental and systemic impacts.
• Exploring safety-aware deployment strategies with emphasis on uncertainty modeling, misuse risk, and constraint-based system design.

BlueDot Impact

• Completed mentored technical training covering alignment theory, threat pathways, interpretability, scalable oversight, and AI risk evaluation.
• Developed a concrete AI control and deployment safety roadmap focused on strengthening preventive and constraining layers in real-world systems.

BlueDot Impact

• Building an Agent-Based Red Teaming Framework — a lightweight evaluation pipeline that automatically tests LLM-powered chatbots and applications for safety vulnerabilities prior to deployment.
• Framework targets four risk categories: prompt injection, data leakage, hallucination, and unsafe tool usage — risks that standard benchmarks fail to surface under real-world adversarial conditions.
• Four-stage pipeline: curated adversarial attack prompts → target LLM endpoint → rule-based safety judge (pattern matching + classification heuristics) → structured safety report with scores, failure list, and triggering prompts.
• Representative output metrics include safety score (0–100), injection success rate, hallucination rate, and total prompts tested — designed to be repeatable and accessible to development teams pre-deployment.
• Research goal: assess whether automated adversarial testing can surface safety risks earlier in the development lifecycle and move evaluation beyond fixed benchmarks toward dynamic misuse simulation.

Apart x Martian Hackathon

• Mechanistic analysis of manipulation vulnerabilities in AI orchestration — identifying, ranking, and visualizing attack vectors against judge and router models.
• Collaborated with the Trajectory Labs community to document findings and propose mitigations.