Hindsight, a bootstrapped AI startup, is challenging the assumption that larger language models are better by unveiling a 3-base Small Language Model (SLM) ensemble classifier under 600M parameters. In competitive benchmarks against AI moderation models from Azure, Google, Mistral, OpenAI, and Meta—including models up to 14x its size—Hindsight achieved a 97.9% success rate in detecting weaponized Coded Language, while competing models struggled to break into double digits. Founder Dean Gebert initially built a rudimentary AI model in 2019, but Facebook cut off API access days after launch, despite a pre-launch list of beta clients that could lead to $7M ARR in the first 12 months. In 2025, after finding existing AI content moderation models inaccurate, he taught himself to build a model. He pivoted from monitoring public figures to flagging dangerous language in large data sets. Without institutional backing, Hindsight used local hardware instead of cloud compute. Except for 4 validation trainings on AWS g5.xlarge instances, the entire training pipeline—164 model variants to date—was run on a 2023 MacBook Pro with Apple's Metal Performance Shaders framework, an M3 Max chip, and 64GB of memory, despite 4–6x longer training times.