Single-model approaches failed on domain-specific construction imagery. Accuracy varied wildly across photo types, and costs scaled linearly with volume.

Built a multi-stage vision pipeline: GPT-4V for high-level scene understanding, Google Cloud Vision for object detection, and Claude for structured reasoning over combined outputs. Each stage feeds the next with confidence scoring.

Production system processing thousands of photos daily with 90%+ accuracy on domain-specific classification tasks. Cost per image dropped 60% vs single-model baseline.

Building an agent that operates independently on a schedule, maintains long-term memory, and coordinates multiple specialized sub-agents without human intervention.

Designed a heartbeat-driven architecture with specialized sub-agents for research, memory review, and habit coaching. Semantic memory uses SQLite with embeddings for vector search. Conversations stored as append-only JSONL for crash safety.

Fully autonomous agent running in production. Conducts independent research, reviews and consolidates memories, and coaches habits — all without prompting.

Standard system prompts tell agents what to do. But autonomous agents face ambiguous decisions where instructions run out. Needed a way to shape judgment, not just behavior.

Developed aspiration-based intent prompts — short identity statements that frame the agent's goals as intrinsic motivation. Tested across Claude, GPT, and Gemini using psychometric methodology (Big Five, Dark Triad). Added a severity clause to prevent over-identification with persona traits.

Measurable personality shifts across all three model families. Intent prompts produced more consistent autonomous behavior than instruction-based prompts. Published as ongoing research, validated in production with Alvis.