Fearing for her life, Lina Guillen went to police in January.
Earlier in the day, her estranged husband had allegedly threatened her, and she was concerned about her own safety, as well as that of her four children.
Her case was plugged into a risk assessment system called VioGén, which police in Spain use to predict how much risk women who report domestic, family or intimate-partner violence might be at.
The algorithm gives individual cases a risk score, which determines what protection police and the courts provide.
It's controversial and at times has underestimated the risk faced, with deadly consequences.
"She wanted to change the locks on the house so she could sleep and live peacefully with her children … she wanted independence," Lina's cousin, Daniel Rios, told the ABC.
The VioGén algorithm gave Lina a medium risk score and as a result, the courts refused to give a restraining order.
Within three weeks, she had been murdered.
"The system that was supposed to protect her failed, and the government bears a great deal of responsibility," Mr Rios said.
"It has no heart, it cannot feel, it cannot know or find out what is going to happen."
The 48-year-old mother of four was killed in February, allegedly strangled by her former partner, who then set fire to the home with their children inside.
They made it out, but Lina didn't.
"The children were crying and screaming in terror for their mother who was there. It was astonishing. They are small, they saw everything," Mr Rios said.
The man has since been charged over the attack, which took place in a town near Málaga, in Spain's southern Andalusia region.
Spain developed and introduced VioGén in 2007 to keep track of more cases of violence in the home and to remove human bias from policing.
It has 35 basic questions that aim to build a picture of an individual's circumstances and from the answers, it then assesses a woman's risk of being harmed.
Since its launch, reoffending has dropped 25 per cent.
Professor Juan José López-Ossorio from Spain's Interior Ministry helped develop the software.
"It is hard to know exactly how many lives are saved, but we do know that our capacity to detect lethal cases has increased over the years," he said.
"Many cases have been prevented because we detected the lethal risk to which women were exposed and implemented many protection measures."
While the system has saved lives, some women have also fallen through the cracks.
A total of 274 women entered into the system have been killed, dozens of whom were deemed low or no risk.
Professor López-Ossorio said it was important the technology continued to be improved.
"Not just the VioGén system, not just the risk assessment tool, but I think all institutions need to improve together to achieve something very important: the fact that risk varies over time," he said.
Police have the power to override the score, but officers have been found to accept the results 95 per cent of the time.
"It is true that police officers with a lot of experience using the tool tend to trust the result of the tool," Professor López-Ossorio said.
"The tool significantly reduces biases that arise from the subjectivity that professionals do have. Therefore, for us, the most important thing is to reduce bias by using both elements: the tool and the professional."
'What the system needs … is more humans'
Gemma Galdon is the founder of Eticas, an organisation that investigates the ethical impact of technology.
She said it was only useful to remove human bias if the algorithm's result accurately reflected a woman's circumstances.
"If 95 to 98 per cent of the time, police are just pressing enter on the output. We saw that that was not really meaningful human intervention," she said.
"Basically, the women were just facing an AI system that they had no awareness or control of."
Her company offered to audit VioGén pro bono for three years, but the Spanish government refused. They launched their own investigation anyway.
"None of the engineers ever had data on what happened to women after a risk score was given," Ms Galdon said.
"So, if a woman's killed after being given a non-existent risk score, they don't know, they don't assess, they don't look at the system again to see what could have gone wrong."
The Eticas probe found that between 2003 and 2021, 71 women killed by their current or former partners had been given a risk score of medium, low or negligible.
Spain is an example of how governments across the globe are shifting towards artificial intelligence and predictive technology to solve social problems.
But nowhere more than Spain is an algorithm so intertwined with how police handle cases of violence against women.
Critics say it is difficult to determine where its recommendations end and human decision-making begins and, in many instances, women were not even aware of the role the software played in their case.
At a recent march protesting violence against women in downtown Madrid, thousands turned out to express their anger over women still being hurt and killed.
Among the attendees was Nessrin Bensaid, a lawyer representing families who have lost loved ones to violence in the home.
She is highly critical of the VioGén system and says migrant women are at particular risk of falling through the cracks.
"It is not possible for an algorithm to determine psychological, physical, or economic abuse, because the questions it generates are too general to be able to identify that psychological violence or even violence that happens continuously over time," she said.
"What the system needs is not more technology, what it needs is more humans: more professionals specialised in violence against women, lawyers, social workers, psychologists, translators and interpreters.
"That is not going to be complemented or replaced by artificial Intelligence", she said.
As the need for better protection collides with technological progress, women are hoping their lives are not solely placed in the hands of machines.
