Questions we’re in the middle of.

When an agent calls a tool that calls another model, spend belongs to a tree rather than a flat list of API calls. Working through how scope should nest, what happens when a run is abandoned partway, and how attribution survives a single event loop interleaving many runs.

• Observability
• Agents
• SDK Design
• contextvars

Most weak answers in a retrieval system trace back to retrieval, not generation. Comparing fixed-size, overlapping and structure-aware splitting on the same corpus and question set, to see how much of the gap a better prompt can actually recover. Early impression is: much less than expected.

• RAG
• Embeddings
• Evaluation
• ChromaDB

Human-in-the-loop systems route on confidence, so a miscalibrated confidence signal quietly breaks the entire safety argument. Looking at whether self-reported certainty, output-distribution signals or agreement across samples give a threshold worth routing on.

• Calibration
• Extraction
• Human-in-the-Loop
• Reliability

Running inference locally removes per-query cost and keeps data in place, but replaces those with a hard constraint: much less context and much less tolerance for a loose prompt. Investigating how far careful retrieval and context packing close the gap against a hosted frontier model.

• Local Inference
• Ollama
• Context Design
• Gemma

NDVI rises every growing season whether or not anything was planted, and the seasonal amplitude is large enough to swamp a multi-year trend. Testing decomposition approaches on irregularly sampled series, since cloud cover means observations arrive whenever they arrive.

• Time Series
• Geospatial
• NDVI
• Signal Processing

Collecting the ways these systems actually break in practice — retrieval that returns plausible but wrong passages, extraction that is confidently malformed, silent truncation, cost that scales with a retry loop nobody noticed. Notes toward a checklist worth running before shipping.

• Reliability
• Evaluation
• Systems Design

Notes on the boundary between work that belongs in a request and work that belongs in a queue — and on the interface consequence, since asynchronous work means the user is looking at a job rather than a result, and pending state has to be represented honestly.

• Queues
• BullMQ
• System Design
• UX