Audrius Šaras · Data & AI Engineer

Three kinds of evidence: what I build on my own time, what I build at work, and what I built during my degrees. Every card opens a full write-up.

Personal research

now building

Hibiki LT ↔ EN: real-time speech translation.

Adapting Kyutai's Hibiki simultaneous translation model to an entirely new language pair: Lithuanian and English. Full pipeline, built from scratch. 44,000 curated parallel pairs, synthetic bilingual stereo speech, a custom fine-tuning fork, and on-device inference with MLX on Apple Silicon. Speech in, translated speech out, no cloud.

fig 1. live waveform, lithuanian phonemes pictured mid-flight. ~12.5 tok/s on an M4.

read the full write-up

Work · KPMG

now building

AI solutions on the Microsoft stack.

The current chapter: demoing, researching and building AI solutions with Microsoft AI Foundry, Copilot Studio and Agent 365, and consulting clients on which of them actually fits their business problem. The hard part is rarely the model. It is the match between tool and problem.

read the full write-up

built · 2025

A data platform for oil & gas.

Co-built and managed a Microsoft Fabric data platform for a large oil & gas client: lakehouse, medallion architecture, metadata-driven pipelines, change data capture. Personal contribution: re-architecting pipeline runs from sequential to parallel, cutting processing time more than 5x.

read the full write-up

built · 2026

Real-time route optimisation.

A real-time vehicle routing platform on Fabric Real-Time Intelligence, built with colleagues: Eventhouse, Kafka, Azure Event Hubs, Eventstream, KQL and live dashboards. The team was selected to present it at Microsoft FabCon 2026 in Atlanta.

read the full write-up

built · 2024

HR analytics at 500+ users.

Built and managed Power BI dashboards for HR analytics, each serving more than 500 users. Then made them survive without me: consulting and teaching citizen developers to maintain them.

read the full write-up

Study

built · 2024

Forecasting SKU sales with machine learning.

Master thesis at Unilever and RSM: enhancing category marketing planning by forecasting SKU sales value share. Gradient Boosting and Random Forest against Prophet over a 12-week horizon, SKU descriptions turned into text embeddings, LASSO-targeted predictors, and SHAP values to open the black box.

fig 2. the pipeline in one line: from point-of-sale noise to explainable forecasts.

read the full write-up

built · 2025

The NHS knee models.

EAISI group project on NHS Digital PROMs data: predicting which knee replacement patients will have poor outcomes, before surgery. Imbalanced classes, precision-recall trade-offs, and decision thresholds a clinician could actually use.

fig 3. interactive. hover to flex. predicted range of motion: excellent.

read the full write-up

built · 2024

EcoNexus: finance meets ESG.

RSM FinTech project: a platform that balances financial and sustainability goals in coffee supply chains using weighted goal programming. Python, Flask and the Gurobi optimizer, with dynamic scenario analysis for supplier selection.

read the full write-up

built · 2024

One more hub, fewer kilometres.

Business Analytics Workshop at RSM: finding the optimal extra hub for a logistics client's hub-and-spoke network. K-Means for candidate locations, then TSP and VRP models in Gurobi and Google OR-Tools to minimise kilometres driven.

read the full write-up

Audrius
Šaras.

The career, plotted.

The projects, documented.

BSc Communication Science, minor in Data Science & AI.

Leading Azure migrations at big-tech scale.

MSc Business Analytics & Management at RSM.

Data analytics where shelves meet spreadsheets.

Data & AI engineering, in production.

Post-master in Data Science & AI at EAISI.

Hibiki LT ↔ EN: real-time speech translation.

AI solutions on the Microsoft stack.

A data platform for oil & gas.

Real-time route optimisation.

HR analytics at 500+ users.

Forecasting SKU sales with machine learning.

The NHS knee models.

EcoNexus: finance meets ESG.

One more hub, fewer kilometres.

The credentials, verified.

Get in touch.

The work

Concepts

What it taught me

Audrius Šaras.

The career, plotted.

The projects, documented.

BSc Communication Science, minor in Data Science & AI.

Leading Azure migrations at big-tech scale.

MSc Business Analytics & Management at RSM.

Data analytics where shelves meet spreadsheets.

Data & AI engineering, in production.

Post-master in Data Science & AI at EAISI.

Hibiki LT ↔ EN: real-time speech translation.

AI solutions on the Microsoft stack.

A data platform for oil & gas.

Real-time route optimisation.

HR analytics at 500+ users.

Forecasting SKU sales with machine learning.

The NHS knee models.

EcoNexus: finance meets ESG.

One more hub, fewer kilometres.

The credentials, verified.

Get in touch.

Audrius
Šaras.