Her hair, once as dark as Naomi’s, had turned silver around the temples and nearly white at the front. She still worked three days a week despite claiming every January that she would reduce to two.
She had become a labor attorney almost by accident and stayed because, as Julian used to say, arguing with institutions was the only hobby she had ever managed to monetize.
The meeting took place at Language Access House.
The nonprofit Naomi had served for years had expanded into a national network.
What began with courtroom and hospital interpreters now supported emergency hotlines, schools, immigration clinics, elder care, disaster response, and public-benefit systems.
Chapter 27
Its headquarters occupied an old brick building in Queens.
Nothing glamorous.
Functional.
Wheelchair accessible.
Children’s drawings taped near reception.
A coffee machine that had been replaced six times and remained terrible.
The conference room, however, was new.
A corporate donor had paid for renovation.
That fact was currently irritating Ava.
Across from her sat Ethan Vale, founder of a translation-technology company called BridgeSpeak.
He wore no tie.
Soft gray sweater.
White sneakers.
A watch whose price Ava could have guessed only because one of her grandsons had once shown her a video about luxury watches and then been subjected to a forty-minute lecture about retirement accounts.
Ethan was not arrogant in the obvious way.
That made the conversation harder.
He listened.
He thanked people.
He had grown up in a bilingual household and genuinely cared about language access.
His company had built real-time interpretation software that could translate more than ninety languages with remarkable speed.
BridgeSpeak wanted to provide Language Access House its platform at almost no cost.
The savings would be enormous.
The nonprofit could redirect millions toward legal services and direct aid.
Ava wanted it to work.
That was why she was suspicious.
Ethan pointed toward the screen.
“Our accuracy in common medical contexts is ninety-eight-point-seven percent.”
Ava looked at the slide.
“According to what benchmark?”
“Clinically reviewed datasets.”
“Reviewed by whom?”
“Certified linguists and physicians.”
“How many?”
Ethan gave the number.
Good answer.
Ava continued.
“What happens below confidence threshold?”
“The system flags.”
“Who sees flag?”
“The provider.”
“In English?”
“Yes.”
“What if provider ignores it?”
“We log.”
“Logging after harm does not prevent harm.”
Ethan smiled patiently.
“No system eliminates human error.”
“I agree.”
“That is why I am asking where you leave room for it.”
He leaned back.
“We are not proposing replacing interpreters in high-risk encounters.”
“What counts high-risk?”
“Emergency consent, surgery, psychiatric assessment, end-of-life decisions—”
“Domestic violence?”
“Yes.”
“Child protective cases?”
“Yes.”
“Labor disputes?”
He paused.
“Possibly.”
Ava noticed.
“Possibly is doing work.”
A younger Language Access House attorney named Meera Patel hid a smile.
Ethan saw.
He smiled too.
“I knew Naomi Carter’s daughter would be difficult.”
The room changed.
Not dramatically.
Ava simply stopped moving.
Ethan realized immediately.
“I’m sorry.”
Ava looked at him.
“Why did you say that?”
He opened his mouth.
Then thought better.
“Because I read your background.”
“That tells you who my mother was.”
“Yes.”
“It does not explain my question.”
“No.”
He nodded.
“You’re right.”
Ava waited.
Ethan continued.
“I used family history to explain professional scrutiny.”
Chapter 28
“Yes.”
“That was lazy.”
“Yes.”
Meera looked down at her notes.
Ethan sighed.
“I walked into that.”
Ava’s expression softened half an inch.
“Better here than after contract.”
They continued.
The pilot began three months later.
BridgeSpeak handled low-risk scheduling, appointment reminders, transportation coordination, and basic administrative calls.
Human interpreters remained for medical decisions, legal proceedings, safety disclosures, and any conversation where user requested one.
At first, it worked beautifully.
Wait times dropped.
Patients received information faster.
A rural clinic that had struggled to find overnight Vietnamese interpreters could at least confirm appointment details instantly.
A school district saved thousands.
An elderly Korean woman used BridgeSpeak to understand medication refill instructions at two in the morning.
Useful.
Then a farmworker named Mateo Alvarez called a wage-theft hotline.
He spoke Mixtec and Spanish.
His Spanish was functional but not strong enough for legal detail.
BridgeSpeak defaulted to Spanish because its confidence detector identified that language first.
Mateo tried to explain that a labor contractor had withheld three weeks of wages and threatened to report several workers if they complained.
The software translated one phrase as:
My supervisor warned us there would be immigration consequences if we caused problems.
What Mateo actually said, according to a later human interpreter, was closer to:
The contractor said he had already given our names to immigration and showed us photographs of our families.
Different.
Very different.
The automated system classified the call:
Wage dispute with possible intimidation.
A human advocate scheduled a follow-up in four days.
If correctly understood, the case would have triggered same-day safety review.
Mateo never made that follow-up.
He disappeared from the labor camp the next morning.
Not kidnapped.
Not murdered.
The truth was less dramatic and still serious.
He fled.
Turned off his phone.
Stayed with relatives in another state.
For twelve days nobody at Language Access House knew whether he was safe.
When Meera found the translation discrepancy, she called Ava.
It was 6:18 in the morning.
Ava answered with:
“Somebody better be bleeding.”
Meera said:
“Not that I know.”
Ava sat up.
“What happened?”
She listened.
Then:
“Disable Mixtec routing.”
“Already recommended.”
“Recommended?”
“BridgeSpeak wants to investigate first.”
Ava threw back blankets.
“Who?”
“Ethan’s team says model may have switched based on user speech.”
“It did.”
“That is exactly why you disable.”
“They say interruption would cut service for three counties.”
Ava put on glasses.
“Then manual backup.”
“We don’t have enough interpreters.”
“Find them.”
“Ava.”
“What?”
“That is what they’re saying.”
Ava stopped.
There it was.
The difficult part.
Not villain.
Scarcity.
If they disabled translation, hundreds of people might lose access for days.
If they kept it, severe details might be mistranslated.
No perfect move.
Ava took a breath.
“Okay.”
Meera waited.
“Keep administrative functions only.”
“No legal intake.”
“No threats.”
“No employment disputes.”
Chapter 29
“Anything containing coercion terms routes human.”
“We need technical change.”
“Then call Ethan.”
“Already awake?”
“He will be.”
Ethan answered on second call.
He sounded exhausted before anyone explained.
Meaning his team had already found it.
“Model overweights dominant-language detection.”
Ava said:
“English.”
“Spanish in this case.”
“You know what I mean.”
“Yes.”
“Why?”
“We trained for conversational continuity.”
“If user switches languages, it avoids interrupting.”
“Who decided continuity mattered more than uncertainty?”