Stance Stability Evaluation Pipeline
Pipeline for measuring whether a language model's expressed positions hold up after argumentative pressure is removed — built for a weekend AI research sprint
Full-stack developer exploring the space between
thoughtful systems and the possibilities of AI.
THE JOURNEY
Projects, people, and everything I’ve built along the way.
Pipeline for measuring whether a language model's expressed positions hold up after argumentative pressure is removed — built for a weekend AI research sprint
AI-powered job search assistant that helps users find relevant job opportunities based on their skills, experience, and preferences. The assistant provides personalized recommendations and streamlines the job application process.
Designed an AI Orchestration Layer managing the full LLM workflow — data preparation, prompt construction, model routing (local vs API), output parsing, and validation guardrails, versioning and with human-in-the-loop — projecting ~33% reduction in clinician documentation time per reporting cycle.
Two-stage classification: Claude API (Haiku) for topic judgment, GPT for format summarization. Each model used where it outperforms the other. Designed to run before data desensitization to prevent placeholder contamination in training sets.
Designed and developed event management system to send timely reminders for important dates and events, allowing admin flexibly grouping up different categories employees based on their roles and locations to customize reminder preferences and ensuring high attendance and engagement. Implemented with queue-based architecture for scalability and reliability
End-to-end pipeline: export conversation data from multiple AI platforms, score, filter, clean, desensitize, convert, and fine-tune. Architecture insight: scoring before desensitization prevents placeholder leakage. Working adapter: local_claude_final2.
Browser extension injecting branch-marking UI into Claude, Gemini, ChatGPT, and Grok simultaneously, enabling precise branch marking, collapsing, and navigation across hundreds of messages. Platform-specific adapters handle DOM differences across four distinct interfaces.
Designed a scheduling system to orchestra resources across 5+ roles and 3 campuses weekly 2000+ and daily 400+ sessions to be auditable and aligned with student attendance and employee PTO and availability; ensure session manageable under flexible constraints; generate reports for resource coordination and capacity planning.
Designed and developed a bank-like PTO system to manage staff time off, allowing customize PTO requests and track PTO in a transparent and auditable manner, ensuring alignment with scheduling and session coordination.
Designed a goal progress tracking system to manage 30k+ goals and multiple years of data, enabling time shift monitoring of progress towards goals, generating insights with visualization of heat maps for clinicians, and providing actionable recommendations for next steps.
Designed and developed internal systems for session scheduling, wellness digitalization, goal progress tracking, batch reporting,and compliance workflows across multiple roles and campuses.6200+ contributions from May 2022 on GitHub
Managed user center and developed KYC verification flow for a cryptocurrency trading platform supporting 68k+ users, ensuring secure and compliant onboarding while maintaining a seamless user experience.
BEHIND THE WORK
Converting complex physical-world logic into high-efficiency digital architecture. Targeting roles that require rigor, scale, and AI integration.
THE NEXT CHAPTER