Picture a customer at an airport counter explaining a problem in Hindi while a handheld device translates for the agent. In thirty seconds the customer has said their full name, their booking reference, perhaps a phone number. In many speech translation systems, all of it travels through the pipeline untouched.
That is the problem at the heart of Redact Before You Translate: Real-Time PII Protection in Multilingual Speech Pipelines, the latest study from engineer and researcher Mallikarjuna Yadav Bachu, published this month. It is his eighth research article since 2019, and it lands as the business world is asking hard questions about what AI systems do with the data people give them.
Bachu’s answer is direct: personal information should be identified and protected at the earliest possible point, before speech is translated, stored or passed to another service. It is a principle that sounds obvious. In practice, few real-time systems are built that way.
Why the timing matters
This year generative AI moved from novelty to boardroom priority almost overnight. Companies are piloting chat assistants, transcription tools and translation services at speed. At the same time, regulators are paying closer attention. The European Parliament adopted its position on the proposed AI Act in June, and data protection rules already require careful handling of personal information.
Speech is a particularly sensitive case. People say things aloud that they would never type into a form, and they rarely think about where that audio goes. When the conversation crosses languages, the data may pass through several models and services in sequence. Each step is a chance for exposure.
Bachu’s study treats protection as part of the pipeline’s design. It explores how names, account numbers, contact details and other identifiers can be detected and redacted in real time, across many languages, without adding delays that would make the conversation feel stilted. The goal is translation that still works for the customer while revealing as little as possible.
One device, many hands
Earlier this year Bachu published a companion study on a related trust problem. Shared Devices, Individual Identities: Secure SSO and Single-Logout Architecture for Multi-User Frontline Mobile Fleets (March 2023) looks at the reality of frontline work: a single handheld often passes between several employees in one shift.
If the device stays signed in as the previous user, data and accountability blur. If every app demands a fresh login, workers lose minutes they do not have. The study sets out an architecture for single sign-on with reliable single logout across many apps, so each worker is properly identified and properly signed out, without slowing anyone down.
Read together, the two 2023 papers make a clear argument. Before AI can be trusted on the frontline, organizations need to know who is using the device and what happens to the data that flows through it.
Five years of research, at a glance
Bachu’s eight studies follow a steady path from operations and devices toward speech, identity and privacy.
| 2019 | Prediction on the ramp
Probabilistic machine learning models that estimate how long bags take to come off one aircraft and reach the next, giving planners confidence ranges rather than single guesses. |
| 2020 | Updates without touching devices
Linking CI/CD pipelines to mobile device management so app updates reach large fleets of rugged handhelds automatically, in controlled waves. |
| 2021 | Visibility and modernization
Real-time RFID tracking on handheld devices in radio-noisy airport environments, and a risk-weighted framework for moving large Xamarin app portfolios to .NET MAUI. |
| 2022 | Breaking the language barrier
A comparison of cascaded and end-to-end speech translation across more than 80 languages, and lightweight speaker attribution so handhelds know who is talking. |
| 2023 | Identity and privacy
Secure sign-on and single logout for shared frontline devices, and real-time redaction of personal information in multilingual speech pipelines. |
From the lab bench to the counter
Bachu’s research tracks closely with his day-to-day work. He is an innovation engineer focused on rapid prototyping of new technology for frontline operations. Over the past year and a half, that has included building voice-to-text and text-to-speech capabilities for more than 80 languages, with automatic language identification and speaker recognition, so agents can hold two-way spoken conversations with customers on their handheld devices. He has also worked on machine learning approaches to baggage routing and dispatch, the operational problem that first drew him into research.
Before that, he spent years as a lead mobile developer and architect, building more than ten enterprise apps with Kotlin, Swift, Xamarin, .NET MAUI, Ionic and Angular. He implemented single sign-on with single logout across a suite of Android apps, set up CI/CD pipelines with Jenkins and device management tools, and integrated barcode scanners and RFID on rugged hardware. Several of his studies grew directly out of problems he met in that work.
What the field can take from it
Three ideas stand out across Bachu’s body of work, and each has value well beyond aviation.
The first is that AI belongs where the work happens. His research consistently puts prediction, translation and tracking on the devices frontline staff already carry, rather than in back-office dashboards they never see.
The second is that uncertainty and limits should be designed in, not hidden. His baggage models report confidence, not just estimates. His translation study looks hard at the languages where systems perform worst. That honesty is what makes a system safe to rely on.
The third, and most timely, is that trust is an engineering problem. Knowing who holds a shared device and protecting what customers say are not policy afterthoughts. They are features that must be built into the architecture from the start.
For hospitals, retailers, transport operators and any organization serving the public in many languages, those lessons apply directly. As AI tools spread across customer-facing roles, the gap between what is possible and what is responsible is becoming the central question for the industry.
The bottom line
Eight studies in, Mallikarjuna Yadav Bachu has built a body of work that connects machine learning, mobile engineering and speech technology through a single concern: making frontline technology that people can depend on. His newest research adds the piece that 2023 has shown to be most urgent. In the race to deploy AI, the systems that last will be the ones that protect the people using them.
