Journey

Research notes, project insights, and ideas in progress.

Time-Series
10 min

Short-Term VIPV Forecasting: Lessons from Two Years of BMW i3 Data

What makes VIPV forecasting fundamentally different from fixed-site solar — and how a causal day-shape prototype framework changed our feature engineering approach.

Time-SeriesEnergy AIVIPV
Read →
LLM & RAG
8 min

When Knowledge Conflicts Are the Point: Building Pingala

Most RAG systems treat conflicting sources as noise. Pingala treats them as signal — exploring what happens when a system must faithfully represent genuinely divergent interpretations.

RAGKnowledge ConflictLLM
Read →
Robotics
7 min

The Gap Between Seeing and Doing: Notes from a Robot Dog Project

Detection confidence and physical execution are different problems. Building a multi-sensor autonomous robot exposed the real cost of that gap — and why SLAM + RL changes the equation.

Embodied AIROS2Sensor Fusion
Read →
Energy AI
9 min

Energy AI: Why Domain Context Changes the Model

Generic time-series models fail on energy data not because they are weak — but because energy has physical structure that only domain knowledge can encode.

Energy AITime-SeriesExplainable AI
Read →
LLM & RAG
11 min

Domain-Specific LLMs: Fine-Tuning vs. RAG — A Practitioner's View

Fine-tuning encodes knowledge into weights; RAG keeps it external and auditable. For high-stakes domains, that architectural distinction matters more than benchmark numbers.

Domain-Specific LLMRAGLLM Evaluation
Read →
Time-Series
9 min

Explainability in Forecasting: SHAP Beyond Feature Importance

SHAP tells you which features mattered — but in time-series forecasting, the more interesting question is when and why operational state dominates raw irradiance.

Explainable AISHAPTime-Series
Read →