hi, my name is mert sonmezer.

i'm an ai researcher -- i build, benchmark, and write about ai models that have to work in the real world.
currently
ai researcher @ neural bridge
bsc cs @ metu -- graduating jun 2026
previously
harvard med school
virginia tech
neural bridge
metu-dtx
Mert Sonmezer

about

i work at the intersection of intersection of research and production systems. my research has focused on retrieval-augmented generation, generative models, and how to adapt frontier models to domains where the data is messy, the stakes are high, and the benchmarks don't quite exist yet -- medicine, finance, etc.

i've spent roughly equal time on both sides of the research-production line. three first-author papers at top-venues; ~2 years building llm agent systems and rag pipelines at neural bridge. i think those two tracks teach each other things, and the goal of most of my work is to make that exchange a little more concrete.

-- bsc in computer engineering @ metu, graduating june 2026.

-- first-author papers at iccv 2025, midl 2026, and neural networks. arxiv: 2510.15022, 2602.07872, 2504.17913.

-- i research how foundation models generalize to high-stakes vertical domains: finance, medicine, and scientific imaging.

-- based in ankara.

experience

-- now
ai researcher
new york, ny · remote
bridging frontier ai research and production. researching tabular, time-series, and finance-oriented foundation models and benchmarking them on real funding tasks.
--
boston, ma
built a retrieval-augmented, region-aware medical clip pipeline for pediatric wrist-fracture detection. resulted in first-author midl 2026 paper. advised by dr. sila kurugol.
--
research intern
blacksburg, va
designed a submodular lora-selection algorithm for diverse text-to-image generation. resulted in first-author iccv 2025 paper. advised by dr. pinar yanardag.
--
ai engineer
new york, ny · remote
built user-programmable llm-based multi-agent systems and an llm-powered synthetic data synthesizer as part of a 10-engineer team.
--
research intern
ankara, turkey
proposed a lightweight cnn architecture and a normalization framework for non-stationary time-series forecasting. resulted in first-author paper in neural networks. advised by dr. seyda ertekin.
--
ai engineer intern
new york, ny · remote
fine-tuned falcon, llama-2, and gpt-4 for context-question-answer tasks. released the neural bridge rago model family and rag datasets on hugging face.

selected work

news

2026
paper accepted -- midl 2026
wristmir · coarse-to-fine retrieval for pediatric wrist x-rays
2026
paper accepted -- neural networks
canet · non-stationary time-series forecasting
2025
paper accepted -- iccv 2025
loraverse · submodular retrieval over 100k+ adapters
2025
most innovative research award -- ceng488 guided research
metu · fall 2024-25
/now
what i'm
up to
last updated · 10 may 2026
  • finishing my final year project at metu -- presenting june 2026.
  • researching foundation models for funding at neural bridge.
  • continuing medical research with a marvellous team at harvard med school.
  • weighing master's offers and what comes next.