AI Systems

Writing filed under AI Systems.

On-device AI fits private workflows

On-device AI brings models closer to the user, which matters for privacy, latency, offline use, and personal tooling.

June 8 20263 min read

#on-device ai #privacy #local models

Smaller models belong in the architecture

A practical note on using smaller models for focused jobs instead of sending every request to the largest model.

December 30 20257 min read

#distillation #small models #training

Recommender systems are taste machines

Recommendations are not just ranking with a nicer label. They are a machine for encoding taste, feedback, and what a product thinks matters.

June 25 20255 min read

#recommenders #ranking #personalization

Threat analytics is time series work too

Security signals often come down to timing, baselines, drift, and whether a pattern belongs to the environment.

May 7 20254 min read

#time series #forecasting #sequence models

o3 and o4-mini made routing harder to dodge

The o-series releases made the cheap, fast, smart tradeoff harder to avoid and better to name.

April 23 20256 min read

#openai #routing #reasoning

Claude 3.7 made thinking a product control

Claude 3.7 Sonnet made hybrid reasoning visible enough that teams had to decide when thinking should cost time.

February 28 20255 min read

#claude #reasoning #latency

Local models are part of the platform

Local inference is useful for privacy, latency, and focused workflows, especially when it is treated as part of the system.

November 19 20242 min read

#local models #small models #privacy

Reasoning models changed where latency belongs

The o1 preview made it easier to talk about spending extra time on hard steps instead of every step.

September 18 20242 min read

#reasoning models #latency #openai

Routing is the quiet part of AI platforms

A short note on model routing, context routing, and why production AI systems need more than one path.

May 17 20246 min read

#attention #transformers #neural networks

Long context changed retrieval expectations

Gemini 1.5 made long-context work feel less theoretical, but it did not remove the need for retrieval judgment.

February 19 20242 min read

#long context #gemini #retrieval