Daniel

Explored architectures for real-time hand detection on phones; improved structure and training under tight latency—about +13% mIoU vs prior baselines.

• Lightweight backbones and MobileNet-class swaps under strict compute budgets
• Tuned the accuracy vs latency trade-off for live use
• Supported multi-class, RAW input, and false-positive reduction for production

Built pipelines to synthesize segmentation training data; studied generalization, domain shift, background bleed, and lighting gaps vs real captures.

• Green-screen matting with foreground/background separation and edge/color artifacts
• Randomized brightness and lighting to diversify synthetic sets
• Measured synthetic vs real gaps and how domain shift affected generalization

Deepened capsule networks for hierarchical features: proposed capsule convolution and capsule pooling to deepen stacks while compressing inter-layer parameters.

• ~+10% accuracy on CIFAR-10 vs the best capsule baseline at the time (single model)
• ~50% fewer inter-layer parameters in pooling with similar accuracy

A knowledge-workflow tool for the AI era: turning linear study paths into structured knowledge networks and lowering the cost of organizing insights over time.

• Ingest PDFs, e-books, web pages, and AI chat into one linkable knowledge space
• Combine reading comprehension with mind-map-style structuring to surface relationships between concepts
• Unify reading, notes, cards, vocabulary, and brainstorming to reduce context switching
• Explores durable knowledge and memory—not just another note app with a chat box

A multi-agent system to “think a project through”: agents play product, tech, business, and user roles—cross-examining, debating, and stress-testing assumptions until the plan is more actionable. Public channels plus private agent-to-agent chats for realistic collaboration dynamics.

• Multi-role AI collaboration that mirrors team debate—not single-shot Q&A
• Multi-round iteration to mature plans and surface edge cases
• Dual-layer structure: public discussion plus private agent side-channels for emergent behavior

Explores AI as a neutral helper in relationships: separate conversations with each side, mapping events and emotional blockers without acting as a messenger or moral judge.

• Bilateral private chats to reduce head-on conflict in heated moments
• Strict neutrality: no taking sides, no moral verdicts, no steering outcomes
• Long-term product angles for mediation, emotional triage, and communication scaffolding