Firat Elbey

elbey-1 / model card

Model card

This is the model card for elbey-1: selected product results, capabilities, and known limitations.

Model type

Human, product leadership class

Checkpoint

Group Product Manager, Data & AI, Google

Training

Biochemistry, bioinformatics research, and more than 10 years in product development

Location

Bellevue, Washington

01 / Results

Selected results

Self-learning

Knowledge Catalog context enrichment I led product definition and the preview launch of capabilities that help agents identify and fill gaps in their context.

Nova 1 & 2

Foundation-model launches on Amazon Bedrock I owned requirements across the portfolio and worked with researchers across pre-training and post-training.

2023

Alexa Let’s Chat preview I defined the conversational experience and its LLM-orchestrated architecture, then led PMs across evaluation and launch planning.

20×

Faster end-to-end MRI processing at IXICO Improvements across the imaging pipeline reduced the time from MRI upload to processed result by a factor of 20.

Months → weeks

Clinical-trial setup on IXI Platform I led changes to configuration and onboarding that reduced the time needed to set up a trial.

20+ engineers

Roche VENTANA DP 200 development at FFEI I led product definition and launch with Roche, Ventana and the engineering team.

02 / HUMAN VS FRONTIER

Model card

This is a playful comparison, not a measured benchmark. The career entries describe my experience; the human and model comparisons are illustrative.

Comparison of elbey-1, elbey-2, and a frontier model across training, knowledge, product deployment, judgment, embodiment, and safety.
Open the full-size image to zoom.

† elbey-2 remains in pre-training. Benchmark evaluations are still pending.

‡ Human input tokens are an order-of-magnitude estimate (mostly papers, PR/FAQs, and caffeine).

03 / Capabilities and known limitations

Capabilities and known limitations

Capabilities

  • At Nova, I set portfolio direction and worked with PMs on individual capabilities, balancing model quality, latency and cost.
  • I developed evaluations and reinforcement learning verifiers that guided Nova’s training on reasoning efficiency, coding and model behaviour.
  • At Google, I lead product for Knowledge Catalog’s ontology generation and develop Open Knowledge Format for sharing knowledge across agents and systems.
  • I have hired, managed and mentored teams across product, data science and engineering. Cross-functional team sizes on this page are not direct-report counts.
  • At LexisNexis, I created Context Expert Analytics for the UK and Australia, connecting case-law data with Elsevier’s publication records.

Known limitations

  • Evaluation plans are required before release.
  • A convincing demo does not replace an evaluation.
  • My own adaptive compute remains a work in progress.
  • British spelling (colour, behaviour) occasionally appears in US documents.