How Upwork helps freelancers win more work with Llama

Upwork, one of the world’s largest work marketplaces, is using Llama to power Uma, its mindful AI, to help freelancers land jobs faster and more confidently. Proposal writing can be a time-consuming process, especially for new or under-marketed freelancers. It demands precision and a deep understanding of how to effectively pitch services. Uma’s intelligent system surfaces key work requirements, suggests relevant job skills, and drafts tailored proposals in real time, helping freelancers improve their success rates in a competitive hiring environment.

Upwork’s AI and machine learning team knew that Uma could improve the freelance experience by helping to streamline workflows, so freelancers can save time and focus on higher-value work. To do so, the team needed to upgrade Uma to ask pointed questions about their skills and draft highly personalized proposals tailored to specific job postings.

“Uma is the brain behind key parts of Upwork’s hiring and matching process,” says Andrew Rabinovich, Head of AI and Machine Learning at Upwork. “Uma’s ability to create tailored proposal drafts helps reduce friction for freelancers by helping them accurately communicate their services and offerings, positioning their unique skill sets and expertise to clients to stand out and win more work.”

By integrating Llama 3.1 70b with low-rank adaptation (LoRA), Upwork built a more context-aware, responsive AI that understands the nuances of its platform.

Upwork initially launched Uma to support businesses on the platform with job post creation and talent matching. However, as usage grew, the company saw an opportunity to flip the lens and offer freelancers the same level of assistance by employing AI. In trials, off-the-shelf AI models performed poorly as they were unaware of Upwork’s internal knowledge base. To succeed, the team needed a model they could securely fine-tune and deploy in a way that was both cost-effective and scalable. With Llama, the team developed a deeply customized solution that fit seamlessly into its workflows.

“Upwork needed to deliver high-quality, instantaneous AI assistance in a complex, multi-model environment that connects freelancers with businesses around the globe,” Rabinovich explains.

With Llama, Uma now operates as a multi-agent AI system that dynamically supports freelancers throughout the proposal process. It extracts job requirements from postings, understands each freelancer’s unique background and past successes, and generates tailored pitches that align their strengths with each opportunity.

The Upwork team fine-tuned Llama 3.1 by creating two custom datasets: one that captured Uma’s conversational tone and personality, and another focused on high-performing proposal writing. These datasets were created by top-rated screenwriters and copywriters from the Upwork platform and then used to generate synthetic training data.

At runtime, the Fireworks AI platform deploys LoRA adapters on demand, integrating seamlessly with Upwork’s infrastructure. The LoRA models allow Upwork to build multiple task-specific expert services for each conversation, such as profile review, job analysis, real-time feedback, and proposal generation. Because Uma’s expert models share a single base on the Llama 3.1 foundation, Uma can support highly customized services at a significant scale. The approach also minimizes per-token inference costs, making it efficient to serve millions of freelancer interactions worldwide in real time.

Since scaling Uma’s capabilities, Upwork has seen a 58.8% increase in engagement compared to previous tools, along with a 24% improvement in positive user sentiment. The shift to open source models has also resulted in a 19.25% reduction in the cost per token compared to proprietary models. Together, these gains indicate a faster and more effective experience for freelancers seeking to stand out and secure more gigs.

Thanks to Llama’s open framework, Upwork can maintain full control over data ingestion pipelines, securely store model weights on-premise, and protect them against depreciation—all while reducing development time and supporting long-term versioning.

Uma’s deployment spans Fireworks AI, Amazon Web Services (AWS), and Upwork’s proprietary infrastructure, which Rabinovich says gives the team flexibility to securely deploy open source models at scale. Compared to proprietary models, Llama also helped Upwork lower compute costs and improve performance on long-context tasks. “Its open source foundation made it possible to develop and deploy faster while ensuring long-term security and adaptability,” Rabinovich says. “Llama’s flexibility and precision have been pivotal in delivering these impactful results.”

As for what’s next, the Upwork team plans to continue expanding Uma’s capabilities across the platform, using the same multi-LoRA framework to develop new models aimed at delivering better work outcomes for freelancer and business workflows.