Model type
Human, product leadership class
01 Training data
Five eras of AI, in order
No era skipped. Each one is a shift in the technology, worked from the inside.
2012
Bench science
Infectious-disease research, with two peer-reviewed publications (Scientific Reports, Microbiology).
2015
Regulated medical AI
Digital pathology scanners and neuroimaging platforms for Parkinson's and Alzheimer's trials, when deep learning was new.
2019
Language models, v1
BERT-based legal NLP at LexisNexis, before transformers were mainstream.
2020
LLM agents
Alexa AI: a knowledge graph growing by a billion facts, then "Let's Chat", precursor to Alexa+ and one of the first tool-using LLM agents at consumer scale.
2023
Foundation models
Amazon Nova, launched on stage at re:Invent 2024 and 2025.
2026
Agent context
Data infrastructure at Google that gives agents the right context for long-horizon tasks.
02 Evals
Benchmarks, each with an anchor
Every number is from a shipped product. Where a figure is hard to picture, it comes with something you can.
Billions/day
Tokens served in production by Amazon Nova 1 and 2
Roughly the full text of English Wikipedia, every day.
10,000s
Enterprise customers building on Nova through Amazon Bedrock
2 → 5+
Average turns per dialog on Alexa's "Let's Chat"
Conversations more than doubled in length because people chose to keep talking.
1B+
Facts added to Alexa's knowledge graph at 95%+ precision
At one fact per second, a person would need 30 years.
70 → 90%+
Entity precision in Alexa's catalog
From three misses in ten to fewer than one.
$10Ms
Trial costs saved in Parkinson's and Alzheimer's drug trials
Automated brain measurement, 20x faster analysis of the scans that show whether a drug works.
03 Interactive eval
OVERTHINK: the routing eval
A state-of-the-art reasoning model once spent 17 seconds on "what is 1 + 1". For the next 8 queries, you are the router: answer fast (1 unit of compute) or think deep (15 units). Thinking is always right. Fast is cheap, and wrong on hard queries.
QUERY 1 / 8
COMPUTE SPENT: 0
> loading…
04 Capabilities & limitations
What it does, and what it won't
Capabilities
- 0-to-1 launches through scaled adoption, consumer and enterprise.
- 10+ PR/FAQs and business cases reviewed at C-suite level.
- Product strategy across multiple SVP orgs at once.
- Still hands-on: prototyped a multi-agent application within 6 weeks of joining Google.
Limitations
- Refuses to ship without an evaluation plan.
- Will ask for your eval set before your demo.
- Documented overthinking problem (Amazon Science, 2025).
- Occasionally emits British spelling.
06 Intended use
Products with real users.
I build foundation model and agent products that have to work in production. For roles, talks, or comparing notes on agents and context infrastructure, find me through the links below.