Work

I work on the SQL Server engine for Microsoft Fabric Data Warehouse and Azure Synapse Analytics.

Profile on LinkedIn

  1. Microsoft

    Belgrade, Serbia

    Software Engineer II

    – Present

    Software Engineer

    –

    I am the primary maintainer for AI and Fabric Functions, collations, and T-SQL scalar UDFs in Fabric Data Warehouse.

    I shipped ON ERROR in public preview for seven T-SQL AI functions and extended it to Fabric Functions. Parser, binder, and runtime changes let callers return a fallback value or stop a query on failure.

    For AI-service authentication, I built token exchange so a user's queries on the same warehouse can reuse a cached token. Production traces showed about 90% fewer fetches than a modeled baseline of one per query per process.

    I cut weekly serverless SQL metadata-sync alerts by over 70% year over year using failure-rate and per-workspace checks. Both the old and new monitors caught a dependency outage in side-by-side testing.

    I validated filter pushdown and Parquet statistics to help ship case-insensitive collations on schedule.

    I also shipped a fixed-memory P-squared estimator with allocation-free updates. Its per-thread P50/P90/P95 summaries report AI-call latency per query for production monitoring and regression checks.

    AI functions documentation Shared metadata documentation Collation documentation Scalar UDF documentation

  2. NextMove

    New York, NY

    Developer

    –

    I built NextMove, an email-first job-discovery tool, and made all its product and engineering decisions.

    People could choose companies for new-job emails and industries for curated news. Email meant no extra app or repeated logins; I later added a web interface.

    I used embeddings to deduplicate news, and I trimmed job descriptions before sending selected content to more capable models for curation and job-candidate matching.

    I refined it with peer feedback, but it had no monthly paying users and I never advertised it. I stopped investing because I expected AI assistants and competitors to erode its differentiation, making distribution the harder problem.

    The platform is no longer maintained.

    Visit NextMove
    NextMove's public homepage, including its introduction and feature overview for job discovery, application tracking, and alerts.

    Public website, September 2026 · Marketing homepage.

  3. Hack4Impact

    Boston, MA

    Technical Lead

    –

    Ocean Blue was a volunteer-coordination app for a beach-cleanup nonprofit. I led the five-developer team building it in React Native and Firebase.

    I spoke directly with the nonprofit, reworked the backlog around an MVP, and guided technical decisions, sprint planning, and code reviews. I handed off the project when I left Hack4Impact.

    Portfolio concept, 2026 · AI-assisted; synthetic events.

    A modernized Ocean Blue portfolio concept showing event discovery and signup, with sample cleanup events.

  4. Microsoft

    Belgrade, Serbia

    Software Engineering Intern

    –

    I built and handed off an event-driven load simulator for Azure Synapse Analytics. The work later evolved into a component that remains in use.

    The simulator started from production configurations and explored alternatives, narrowing the candidates for production testing. The live system remained the final check.

    I built telemetry parsing, workload generation, simulation, and analysis in C# with .NET dependency injection, using KQL and Python for reporting.

    Telemetry parsing produces event records; workload generation turns them into a workload specification; an event-driven simulator produces simulated signals; analysis compares them with separate reference measurements to produce a report.
    Telemetry parsing produces event records; workload generation turns them into a workload specification; an event-driven simulator produces simulated signals; analysis compares them with separate reference measurements to produce a report.

    Generalized reconstruction · Internal records and measured traces omitted.

  5. Boston University

    Boston, MA

    Research Assistant

    –

    I mapped and validated overseas development-finance project data using geolocation analysis at the Global Development Policy Center.

  6. Endava

    Belgrade, Serbia

    Back End Development Intern

    –

    I built an internal project-management application with NestJS, TypeScript, and MongoDB, including authentication, REST APIs, and Jest test coverage.

Research

Text to Graphics by Program Synthesis with Error Correction

Co-author, Boston University · 2023

Generative Models for Computer Vision workshop, CVPR 2023

We used language models to generate and edit Python and JavaScript programs for 3D scenes, using graphics and simulation libraries rather than generating pixels directly.

Generated programs run in tools such as Blender and three.js. When execution fails, the workflow feeds the error back to the model for repair; follow-up prompts can change an existing scene.

A program can render successfully and still miss the user's intent. The paper evaluates execution and intent separately, using tasks such as object counts, procedural geometry, and changes to a simulation's gravity.

Read the paper (PDF) Workshop listing
BeforeBefore the instruction: the rendered scene contains a small reflective sphere.AfterAfter the code edit: the same scene contains a larger sphere with visible reflections.

PromptIncrease the size of the ball in the following javascript scene

Added lineOOI.sphere.scale.set(2, 2, 2);

Figure 1, Nikitovic et al. (2023) · One example: GPT-4 edits an existing three.js scene. Before and after use the same crop and scale.

Education

  1. Cornell Tech

    Master of Engineering, Computer Science –

    Cornell University

  2. Boston University

    Bachelor's degree, Computer Science –