How Slingshot AI Accelerated Mental Health AI with Fine-tuning at Together AI

About Slingshot AI

Slingshot AI builds the first foundation model specifically designed for psychology. Founded by Neil Parikh (Casper co-founder) and Daniel Cahn (AI researcher), the company addresses the global mental health accessibility crisis through specialized AI.

Slingshot has developed the largest dataset of its kind of clinically relevant data that enables models to learn from real therapeutic practice, understanding individual contexts, cultural backgrounds, and relationship dynamics that make therapy effective. Unlike generic AI assistants trained to be helpful, Slingshot's models are trained to be therapeutic, knowing when to push back, when to stay silent, and when to offer new perspectives.

The Slingshot team operates a sophisticated multi-stage training pipeline, using Together AI for key components, such as supervised fine-tuning with clinical conversations and Direct Preference Optimization for behavioral policies.

The Together AI Solution

Slingshot evaluated major cloud providers and ML platforms and chose Together AI for three capabilities that standard platforms couldn't match.

Together AI's platform was the most convenient solution supporting frequent fine-tuning cycles through programmatic API calls. Other platforms couldn't offer the same level of reliability or support advanced features that were required by Slingshot to produce their best models.

"The technical challenge was running our multi-stage pipeline reliably at the conversation lengths our therapy models require," explains Daniel Cahn. "Together's platform eliminated the context length constraints and job failures we hit elsewhere, letting us experiment rapidly." 

This enabled Slingshot to run a custom multi-step Supervised Fine-Tuning (SFT) and Direct Preference Optimization (DPO) pipeline, processing extended conversation contexts while orchestrating parallel training runs across multiple experiment configurations and a broad range of open foundation models.

Moreover, Together eliminated the need for GPU infrastructure management for these critical steps in their training pipeline that would otherwise require significant additional engineering effort. "The comparison for us is just to bare metal, which is basically just a matter of headcount," explains the team. "There's a lot that we can build, but by the time we build it, we know Together will already be off building the next thing we’ll need." This allows Slingshot to focus more of their team on model development and their core application rather than training framework improvements and large-scale experiment orchestration.

Lastly, unlike vendor relationships, Together AI operates as an engineering extension through shared Slack channels with daily communication and joint development. "When we identify a need, Together typically delivers within days," notes the Slingshot team. This focused partnership lets Slingshot maintain their own infrastructure where needed while leveraging Together's expertise for compute-intensive training cycles.

"Slingshot is genuinely on the bleeding edge of what's possible with AI," our engineering team says. "Working with them pushes our platform forward: they focus on groundbreaking model development while trusting us to solve the infrastructure challenges."

Business Impact

Infrastructure Efficiency: Together Fine-Tuning Platform eliminates the complexity of managing training infrastructure, allowing Slingshot's team to focus entirely on improving their product rather than ML job orchestration and training reliability.

3x More Frequent Clinical Iteration: Training 3-7 times per week (vs. weekly previously) enables rapid incorporation of clinical feedback. "The day we started our first DPO training run, we just saw a huge stepwise improvement," explains the team. Clinical insights reach production models much faster than before. This speed matters in mental health: better models mean better support for people who need it.

5x Cost Advantage vs. Closed-Source: Open-source models with Together's fine-tuning achieved superior performance compared to five times more expensive closed-source alternatives while providing complete behavioral control.

Pioneering New Techniques: As Slingshot pushes into other cutting-edge methods, the partnership continues to evolve, with both teams learning from each other.

"For us, it's all about speed. We're continuously shipping new models, deploying them, seeing what happens. The faster the pace of iteration, the better we think our models will get over time." — Daniel Cahn, Co-founder & CEO, Slingshot AI

"We take the responsibility seriously," notes Neil. "Everybody here is mission-oriented because we really care about this. Together AI's infrastructure enables us to focus on what matters most: building AI that genuinely helps people."

Get Started

Together AI is proud to support Slingshot's mission by offering a training platform that enables AI developers to build world-leading models.

‍Get started with fine-tuning at Together now and create custom, more efficient models for your tasks. If you would like to explore custom solutions for model training or have some feature requests, contact our team!