Philip Abao & Soraya Johnson · USA, Japan
§ 01
Research notes
- No. 04//Evaluation/6 min
Project Pigeon: a small model that holds long context
Results from three internal checkpoints. The 865M v7 stays within a few points of the much larger v5 on most shared benchmarks, leads on Winogrande, and passes a long-context retrieval test that both larger checkpoints fail.
Read note - 03Mar 28, 2026EvaluationAdversarial robustness in domain-specific models: red-teaming beyond the generic benchmark.Mar 28, 2026 · Evaluation · 7 min7 min
- 02Feb 27, 2026ArchitectureInterlocking specialized models: routing and merging domain experts for compound AI systems.Feb 27, 2026 · Architecture · 8 min8 min
- 01Feb 24, 2026MethodsImproving synthetic data generation bounds via constrained decoding.Feb 24, 2026 · Methods · 7 min7 min
Fig. 02 — neognathae.com
§ 03
Milestones
- 2025.Q4Domain annotation serviceComplete
- 2026.Q1Accepted to NVIDIA Innovation LabComplete
- 2026.Q2Post-transformer researchActive
- 2026.Q2Kestrel, a text classifierReleased