On-Device / Edge AI
Accepted
Efficient and Interpretable On-Device Polynomial Inference for Multi-Variable Ocean Profiling
R² 0.973 ocean temperatureR² 0.895 / 0.794 dissolved oxygen / salinity0.88 ms inference on ESP32 (180 µJ)1.12 ms inference on RP2350 (96 µJ)4.4 µs desktop inference
Interpretable polynomial zone ensembles reconstruct global ocean temperature, salinity and dissolved oxygen, and run on flight-class microcontrollers with no cloud dependency and a sub-megabyte footprint.
IEEE eICocean 2027 (Qingdao)
Remote Sensing / SAR + TinyML
Accepted
Deterministic and Auditable On-Device Polynomial Inference for On-Board SAR Oil-Spill Screening
G-mean 0.822 Deep-SAR SOS benchmark657 → 82 KB int8 model (G-mean cost 0.001)15 KB model on ATmega328P (G-mean 0.803)3.76–9.2 ms per tile on 32-bit microcontrollers54,371 held-out tiles
A frozen PCA front end and closed-form degree-2 polynomial head are folded and quantised into a fixed-point kernel, screening SAR tiles (Sentinel-1 C-band and ALOS PALSAR L-band) for oil spills entirely on-board with a deterministic, auditable model.
IEEE EIECS 2026
Edge AI / Physics
Submitted
Real-Time Phase Classification and Instability Prediction in Bose–Einstein Condensates with Edge-Deployed Interpretable Polynomial Surrogates
27.8 µs on RP2350 (Cortex-M33)33.4 µs on ESP32300 B int16 model
An interpretable polynomial surrogate classifies Bose–Einstein condensate phase and predicts supercritical collapse in real time on microcontrollers.
IEEE ICET 2027 (Chengdu)
Compilers / HPC
Accepted
LEAP: Lifting Legacy x86-64 Binaries to Recompilable IR for Automatic Parallelization
958× peak GPU offload speed-up10.4× geometric mean over -O02.2× vs. -O2 baseline98–100% liftability (43 programs)377 tests pass
Lifts legacy x86-64 binaries to recompilable LLVM IR and automatically parallelizes them across CPU and GPU — no source code, annotations, or recompilation of the original program required.
ITCA 2026 (Guangzhou)
Explainable ML / Medicine
Accepted
Automated Multi-Class ECG Arrhythmia Classification Using an Ensemble of Polynomial Regression Models
78.32% balanced accuracy82.75% six-class accuracy99.7% smaller (123.1M → 330K terms)~1 MB sparsified model
A one-vs-rest ensemble of polynomial regressors classifies six arrhythmia classes from ECG images, with a sparsified model 99.7% smaller within ~1 percentage point of the full ensemble.
IEEE ICCA 2026 (Cairo); journal expansion under review (Wiley)
Explainable ML / Materials
Accepted
Accuracy and Interpretability from a Single Solver: Polynomial Regression up to Degree 5 for Concrete Compressive Strength Prediction
R² 0.859 degree-5 accuracy6.25 MPa RMSE88% fewer terms74% retained terms match concrete physics
A single closed-form polynomial solver (up to degree 5) predicts concrete compressive strength, matching black-box baselines while exposing physically meaningful terms — no gradient descent.
ICACTE 2026
Explainable ML / Design
Submitted
Physics-Informed Polynomial Regression for Certified Concrete Mix Design
R² 0.919 degree-3, 5-fold CV4.73 MPa RMSEgap 0 certified optimality (all instances)16.7% embodied-CO₂ saving12.7–15.6% cost saving vs. measured mixes
Inverts a physics-informed polynomial surrogate via interval branch-and-bound to recommend cost- and carbon-optimal concrete mixes with a certified optimality gap.
ICACTE 2027
Books & Teaching
Published
Books: three titles
3 books 2011, 2022, 20242 editions Programming via OS
Authored three books: Programming via Operating Systems (2022), its Second Edition (2024), and Basics of Programming in OS Linux Environment (2011). Second Edition on ResearchGate. Basics on ResearchGate.
Hilfmann Press · MIPT · 2011, 2022, 2024
Education & Community
Founded
Founder & Lecturer, MutEx School
2019–2023 founded & ledCourses public lectures & mentoring
Founded and ran a programming school in Yerevan, delivering courses and public lectures and mentoring students entering computer science. mutexschool.com.
Yerevan, Armenia · 2019–2023