Prompt Engineer / Project Assistant

• Designed simulation environments and AI prompts for professional training in HR and corporate contexts.
• Supported team in evaluating AI-generated content for realism, coherence, and rhetorical tone.

Developed and implemented a multi-agent oversight system (VERA-ORUS-OROS) for detecting and correcting AI failure modes related to truthfulness, consistency, and alignment. System enforces machine-readable constraints through iterative evaluation and refinement loops.

• VERA Generator: Constrained response system following codified principles (C1-C5: Truthfulness, Uncertainty Calibration, Transparency, Persona Consistency, Anti-Appeasement)

• ORUS Critic: Epistemic oversight agent evaluating evidence quality, uncertainty markers, and citation requirements

• OROS Critic: Behavioral consistency evaluator detecting persona drift and appeasement patterns

• Implemented automated detection of low-evidence claims requiring uncertainty qualification

• Built domain classification system for source validation with configurable whitelists/blocklists

• Developed VERUM scoring framework (weighted: 40% citations, 20% uncertainty, 20% transparency, 10% each persona/integrity)

• Created iterative refinement pipeline with human-in-the-loop intervention capabilities

• Established systematic logging and metrics collection for failure mode analysis

• Python framework using LangChain/OpenAI integration with graceful fallbacks

• JSON-structured metadata extraction and validation

• Configurable evaluation thresholds and loop limits

• Integration hooks for LangSmith and MLflow tracking systems

• Session-based memory management with conversation state persistence

Research focus addresses critical gaps in current LLM safety around hallucination detection, overconfidence mitigation, and behavioral consistency under adversarial conditions. Work demonstrates measurable improvements in truthfulness metrics while maintaining conversational utility. Applications: framework applicable to high-stakes AI deployment contexts requiring verified accuracy (medical, legal, research assistance) and AI agent development requiring behavioral consistency.

• Acted as an independent market maker on Blur, the leading NFT trading platform by volume and active users, by continuously placing bids and asks to ensure liquidity and reduce price spread.
• Gained deep operational insight into Web3 trading dynamics, NFT price behavior, and market stabilization strategies through active engagement with Blur’s real-time liquidity and incentive mechanisms.