› Protime Labs

Research and applied work.

The research streams Protime is actively working on, distinct from the consulting work on Work and the essays on Writing. Public summaries here; deeper artifacts are available on request.

Edge computeActive research

Shadow AI detection on the edge, complementing Purview DSPM.

Technologies
NVIDIA JetsonEdge inferenceNetwork traffic analysisPattern matchingAPI credential detectionPurview DSPM (complementary)SMB deployment

Enterprise Microsoft Purview DSPM works for organizations large enough to operate it. SMB environments inherit the same shadow AI risk without the same governance budget. This research builds an edge-compute device that watches network traffic for AI service indicators and exfiltration patterns — the missing layer between endpoint controls and cloud-side data governance.

Read the writeup
AI alignmentWorking draft · seeking academic sponsorship

Intrinsic Thought Process — alignment at the cognitive level, not the constraint level.

Technologies
DreamerV3 world modelReinforcement learningReward shapingWorld model game theoryMinecraft experimental sandboxAcademic white-paper format

Most alignment work is constraint-based — guardrails, RLHF preference tuning, output filtering. ITP starts from a different premise: values shape what an agent considers, not just what it's permitted to say. The architecture embeds value structures into the world model and reward function so alignment happens during the agent's reasoning, not after it.

Read the writeup
LLMOps · methodologyApplied across engagements · framework iterating

LLMOps integration framework for the enterprise.

Technologies
Anthropic Claude APIOpenAI APIGoogle Vertex AIRAGPrompt engineeringFew-shot learningPrompt injection defenseLLMOps governance

Most enterprise LLM adoption fails not at the model layer but at the integration layer — the workflows, the governance, the operational standards, the Shadow IT pattern that emerges when the central plan moves slower than the user demand. This framework is the operational shape that catches those failure modes early and makes the integration work survivable at enterprise scale.

Read the writeup