# Helping Upwork freelancers win more business with Llama

+58.8%  
higher engagement than previous solutions for freelancers using Uma for proposal development  
+24%  
better sentiment for users leveraging Uma’s Llama model to write cover letters compared to competing models  
-19.25%  
reduction in cost per token compared to proprietary solutions  
15%  
jump in patent filings from accelerated innovation  
*All results are self-reported and not identifiably repeatable. Generally expected individual results will differ.*

## At a glance  
**Industry:** Technology  
**Use case:** Enhancing proposal drafting for freelancers  
**Goal:** Enable freelancers to efficiently create compelling proposals  
**Llama version:** Llama 3.1 with low-rank adaptation (LoRA)  
**Deployment:** Fireworks AI, Amazon Web Services and Upwork’s proprietary infrastructure

[Download full case study](/content/static-resource/llama-case-study-upwork/index.html)

## THEIR STORY  
**Connecting talent with opportunities around the world**  
Upwork is the world’s largest work marketplace, connecting businesses with independent talent across over 10,000 skill categories and 180+ countries. By helping clients and freelancers collaborate seamlessly, Upwork delivers innovative, scalable workforce solutions.

## THEIR GOAL  
**Help freelancers stand out in a competitive marketplace**  
Upwork freelancers, especially people new to the platform or marketing their skills, struggled to stand out and book work. Even seasoned pros found responding to job postings and writing proposals a time-consuming process that required a deep understanding of job qualifications and the ability to convey personal qualities that made them a unique fit for the assignment. Upwork knew that its mindful AI companion, Uma, could help freelancers craft effective, compelling proposals faster and improve their win rates. To succeed, Uma would need to deliver high-quality, instantaneous AI assistance in a complex environment that connects freelancers with businesses around the globe.

## THEIR SOLUTION  
**A proposal-writing companion that asks pointed questions**  
Upwork upgraded Uma with a fine-tuned Llama 3.1 model, giving the AI companion the ability to analyze job postings and freelancer profiles, converse candidly with freelancers about their skills and draft effective proposals — all in real-time.

## THEIR APPROACH  
**Multiple expert models provide comprehensive assistance**  
Upwork fine-tuned a general-purpose Llama model for Uma to derive multiple expert models for the diverse roles Uma had to fill. These included tasks such as analyzing complex job descriptions, developing proposals and providing real-time feedback to freelancers. Rather than retraining the entire Llama model for each specific task, the Upwork team used LoRA. With LoRA, only a small fraction of the weights in a model are fine-tuned to achieve expertise in a task. The Upwork team created LoRA adaptations for each use case, resulting in multiple expert models with far less training time. LoRA adaptations load at run time and share a single base Llama model, so the solution can scale up with minimal per-token computing costs. The Upwork team used Fireworks AI to serve up Llama 3.1 and LoRA adaptations on demand, allowing Uma to deliver responsive, expert assistance in real-time.

## THEIR SUCCESS  
**AI-powered proposal drafting transforms the freelancer experience**  
As Uma scaled up to meet the demands of millions of freelancers, engagement grew by 56% over previous incarnations of the AI companion. By simplifying the proposal drafting process and providing expert advice, Uma saved freelancers time and boosted their win rates, proving Llama was the ideal choice for Upwork’s AI efforts.  
- **+55.8%** higher engagement for freelancers using Uma for proposal development versus previous solutions  
- **+24%** increase in first-time-right chip manufacturing  
- **50%** better sentiment for users leveraging Uma’s Llama model to write cover letters compared to competing models  
- **-19.25%** reduction in cost per token compared to proprietary solutions  
*All results are self-reported and not identifiably repeatable. Generally expected individual results will differ.*