Martin Asenov

Work Experience

🇬🇧 Parexel AI Labs

Oct 2025 – Present

Senior Machine Learning Engineer I

Document understanding and generation grounded in verifiable rewards — OCR, vision-language models and retrieval.

Reinforcement learning for document generation. Recast office document generation with verifiable rewards — executable checks replacing an LLM-as-judge — and used them to build a token-efficient OOXML library that beats the OpenAI and Anthropic document skills at ~half the tokens while scoring 40% higher on reward.
Retrieval & document understanding. Built OCR / VLM / retrieval pipelines for large-scale enterprise document workflows, and showed BM25 rivals multimodal retrievers in visually-rich RAG [ICLR Workshop], when OCR beats VLMs, and vice versa (DISCO) [ICLR Workshop], plus corpus-level QA and failure attribution.

🇬🇧 Huawei R&D UK

Aug 2022 – Sep 2025

Senior Machine Learning Engineer — Systems Infrastructure Research

ML systems research across LLM inference, time-series foundation models and accelerator-aware deployment.

Time-series foundation models. Co-developed DAM [ICLR], a transformer-based forecasting model, with follow-on work on lightweight online adaptation [ICML], zero-shot evaluation [ICLR Workshop], fine-grained FaaS forecasting [EuroMLSys], data augmentation [EuroMLSys] and a US patent; deployed production forecasting models that cut infrastructure costs by thousands of pounds per month.
Cloud systems & LLM inference. Published on serverless cold-start analysis [EuroSys], long-term workload characterization [SoCC] and speculative LLM decoding [ICML Workshop].
Heterogeneous accelerator scheduling. Led a seven-person team building a system that predicts runtime and power for unseen ML workloads across truly heterogeneous hardware — NVIDIA GPUs, Huawei Ascend NPUs and x86/Arm CPUs — at 6–10% error, feeding a power-aware Kubernetes scheduler that cuts cluster-wide energy and cost; supervised intern work on RL-based cluster scheduling [IC2E].

🇬🇧 Hyperscience

Apr 2022 – Jul 2022

Machine Learning Engineer

Document understanding — orientation correction.

Improved rotation correction of documents with data augmentation; shipped an MVP with 15% accuracy improvement within two months of starting.

🇦🇹 Amazon Prime Air

Oct 2018 – Dec 2018

Applied Scientist Intern

Computer vision for autonomous delivery drones.

Image segmentation from limited data using domain adaptation.

🇬🇧 DSTL (UK MoD)

Oct 2017 – Oct 2018

Research Scientist

Active sensing for atmospheric tasks with UAVs.

Gas localization and mapping with a UAV using physically-grounded priors over wind and dispersion.

🇺🇸 Thales

Jun 2015 – Aug 2015

Software Integration & Characterization Intern

No-code automation for in-flight entertainment.

Extended Google Blockly into a no-code test automation framework for in-flight entertainment systems.

Education

🇬🇧 University of Edinburgh & Heriot-Watt

Sep 2017 – Jan 2022

PhD in Robotics & Autonomous Systems

RAD group — physics-informed active sensing for highly dynamical tasks.

Thesis: “Architectures for Online Simulation-Based Inference Applied to Robot Motion Planning”.
Supervised by Prof. Subramanian Ramamoorthy and Dr. Kartic Subr.
Gas localization with a UAV, inference of physical parameters from video, and intuitive-physics priors for control.

🇬🇧 University of Edinburgh & Heriot-Watt

Sep 2016 – Aug 2017

MRes in Robotics & Autonomous Systems

Active sensing of spatiotemporal phenomena with UAVs.

Thesis: “Active Sensing of Spatiotemporal Phenomena with a UAV”.
Supervised by Prof. Subramanian Ramamoorthy.

🇺🇸 University of California, Irvine

Sep 2014 – Jun 2015

International Exchange Student — Computer Science

Year abroad as part of the Edinburgh BSc.

Favorite courses: Neural Networks & Deep Learning, Image Understanding, Computer Photography & Vision.

🇬🇧 University of Edinburgh

Sep 2012 – May 2016

BSc (Hons) Artificial Intelligence & Computer Science

Computational neuroscience honours thesis on visual cortex dynamics.

Thesis: “Dynamic Model of Interactions Between Orientation Selective Neurons in Primary Visual Cortex”.
Supervised by Dr. Mark van Rossum.

Publications

Document Intelligence 2026
Multilingual / visually-rich RAG, OCR & VLM benchmarking and failure attribution. ICLR Workshop
Time-series 2023 – 2025
Foundation models for time-series — DAM, online adaptation, zero-shot evaluation, FoldFormer / TSMix and a US patent. ICML · ICLR · ICLR Workshop · EuroMLSys · US Patent
Cloud Optimization 2023 – 2025
Serverless cold starts, long-term FaaS workload characterization, RL cluster scheduling and speculative LLM decoding. EuroSys · SoCC · IC2E · ICML Workshop
Robotics & Sensing 2019 – 2024
Active sensing UAVs, vision-based dynamics, X-ray scattering inversion and medical-robotics imaging. IEEE T-MRB · L4DC · NeurIPS Workshop · Optics Express · CLEO · EJSO · ASBrS · RA-L

ICLR Workshop 2026
Retrieval or Representation? Reassessing Benchmark Gaps in Multilingual and Visually Rich RAG. M Asenov, K Benkirane, D Goldwater, A Ghodsi.
ICLR Workshop 2026
DISCO: Document Intelligence Suite for COmparative Evaluation. K Benkirane, M Asenov, D Goldwater, A Ghodsi.
US Patent App. 2025
Time series data prediction method and apparatus, and storage medium. L Darlow, A Joosen, M Asenov, A Barker, J Wang, Q Deng. (App. 19/316,033)
ICML Workshop 2025
Higher Acceptance Rates for Speculative Decoding with Randomised Drafting. W Toner, M Asenov, R Singh, A Joosen. (Tiny Titans)
EuroSys 2025
Serverless cold starts and where to find them. A Joosen, A Hassan, M Asenov, R Singh, L Darlow, J Wang, Q Deng, et al.
ICML 2025
Lightweight online adaption for time series foundation model forecasts. TL Lee, W Toner, R Singh, A Joosen, M Asenov.
ICLR Workshop 2025
Performance of zero-shot time series foundation models on cloud data. W Toner, TL Lee, A Joosen, R Singh, M Asenov. (I Can't Believe It's Not Better)
IEEE T-MRB 2024
Achieving dexterous bidirectional interaction in uncertain conditions for medical robotics. C Tiseo, Q Rouxel, M Asenov, KK Babarahmati, S Ramamoorthy, Z Li, et al.
ICLR 2024
DAM: Towards a foundation model for time series forecasting. L Darlow, Q Deng, A Hassan, M Asenov, R Singh, A Joosen, A Barker, et al.
SoCC 2023
How does it function? Characterizing long-term trends in production serverless workloads. A Joosen, A Hassan, M Asenov, R Singh, L Darlow, J Wang, A Barker.
IC2E 2023
Learning to Score: Tuning Cluster Schedulers through Reinforcement Learning. M Asenov, Q Deng, G Yeung, A Barker.
EuroMLSys 2023
FoldFormer: sequence folding and seasonal attention for fine-grained long-term FaaS forecasting. LN Darlow, A Joosen, M Asenov, Q Deng, J Wang, A Barker.
EuroMLSys 2023
TSMix: time series data augmentation by mixing sources. LN Darlow, A Joosen, M Asenov, Q Deng, J Wang, A Barker.
PhD Thesis 2022
Architectures for online simulation-based inference applied to robot motion planning. MA Asenov. (The University of Edinburgh)
CLEO 2022
Line-scan Confocal Endomicroscopy for Rapid Digital Histology of Early Breast Cancer. K Vyas, A Ezzat, M Asenov, M Chauhan, S Ramamoorthy, A Jha, D Leff.
EJSO 2022
Preliminary findings of confocal laser endomicroscopy and Raman spectroscopy in human breast tissue characterisation. A Ezzat, K Vyas, M Chauhan, M Asenov, A Silvanto, A Jha, et al.
ASBrS 2022
Portable confocal endomicroscopy for ductal feature characterization: toward margin assessment in breast-conserving surgery. A Ezzat, K Vyas, M Asenov, M Chauhan, A Jha, S Ramamoorthy, D Leff.
L4DC 2022
Vision-based system identification and 3D keypoint discovery using dynamics constraints. M Jaques, M Asenov, M Burke, T Hospedales.
NeurIPS ML4PS 2020
Inversion of Ultrafast X-ray Scattering with Dynamics Constraints. M Asenov, N Zotev, S Ramamoorthy, A Kirrander.
RA-L 2019
Vid2Param: Modeling of Dynamics Parameters From Video. M Asenov, M Burke, D Angelov, T Davchev, K Subr, S Ramamoorthy.
RA-L 2019
Active localization of gas leaks using fluid simulation. M Asenov, M Rutkauskas, D Reid, K Subr, S Ramamoorthy.
Optics Express 2019
Autonomous multi-species environmental gas sensing using drone-based Fourier-transform infrared spectroscopy. M Rutkauskas, M Asenov, S Ramamoorthy, DT Reid.
CLEO 2019
Multi-Species Environmental Gas Sensing Using Drone-Based Fourier-Transform Infrared Spectroscopy. M Rutkauskas, M Asenov, S Ramamoorthy, DT Reid. (AM1K.6)

> open scholar

Technical Skills

Languages

Python, CUDA, TypeScript

ML & Inference

PyTorch, vLLM, Transformers, Time-Series Foundation Models, RAG, VLMs, OCR

Cloud & Systems

Kubernetes, Container Runtimes, Distributed Training, GPU/Accelerator Scheduling

Research

Active Sensing, Bayesian Optimization, Physics-informed ML, Robotics

Non-exhaustive — generally quick and eager to pick up new languages and frameworks where useful.

Interests

Outside work — running, hiking, climbing, cooking, and the occasional late-night paper-reading binge.