Natalie Buchbinder

Profile

I build production AI and data systems, handling everything from real-time ingestion and modeling to deployment and monitoring. Most recently, I've focused on high stakes environments developing operational digital twins, predictive pipelines, and real time forecasting tools using streaming machine data.

My background combines reliability engineering with full lifecycle software development. I look for roles where I can own complex data systems from initial prototype through to production.

Selected engineering work

Kaful — founded, designed & built end-to-end

Production platform — deployed & multi-tenant

Kaful — Streaming CNC Digital Twin

A deployed platform for monitoring CNC cutting tools in real time. After each cut, the system ingests new sensor data, updates its estimate of tool wear, and forecasts how much useful life remains. Rather than returning a single prediction, it produces a range that reflects uncertainty and becomes more precise as more data arrives. I built the platform end to end, including authentication, customer data isolation, streaming ingestion, the wear and RUL models, storage, testing, and deployment.

Designed the system to support multiple machine shops while keeping each customer's machines and data fully isolated. A shared storage interface allows it to run on SQLite locally and Postgres in production, with parity tests across both databases.
Improved the ingest path for scale and concurrency by replacing an operation that slowed as a run grew with a constant time lookup, and adding per run locking to prevent simultaneous uploads from corrupting state. Benchmarked the system before and after the changes.
Deployed on Render, Neon Postgres, and Cloudflare R2, with approximately 136 automated tests.

Python · FastAPI · Postgres · streaming ingest · multi-tenancy & auth · Monte Carlo · Docker

See it live →GitHub kaful.ai

Agentic LLM pipeline — documentation to executable model

Kaful — Document to Twin Pipeline

The engine behind the platform: a seven-stage agentic pipeline that reads raw OEM PDFs, extracts component specs via RAG + LLM, generates a physics-based degradation simulation, and propagates it forward to a calibrated RUL range per component — producing forecasts even when historical failure data is limited or unavailable.

Built and published digital twin demos for four real machines — a CNC machining center spindle, an Atlas Copco air compressor, a Tempress LPCVD semiconductor furnace, and an Eversys espresso machine — directly from raw OEM PDF documentation.
Presented at IWSM 2026 (International Workshop on Software Measurement), American University.
Diagnosed why RUL models fail to transfer across tools: the break is in the wear-to-force observation map, not the degradation parameters — joint state-parameter estimation is unidentifiable from force alone. Shipped a mixture-over-references deployment posture instead of overclaiming transfer.

RAG · LLM extraction · code generation · physics simulation · Monte Carlo · Python

Try the live demo →GitHub

Modeling & evaluation

EPFL & Yeshiva University

Survival Analysis Simulation Benchmark

Reproducible simulation framework benchmarking four survival models — CoxPH, Random Survival Forest, Gradient Boosting Survival Analysis, and a custom PyTorch DeepSurv implementation — under increasing right censoring. 100 trials × 3 censoring levels, evaluated with Harrell's C, Uno's C, time dependent AUC, and Integrated Brier Score.

DeepSurv (Cox partial likelihood loss, Breslow baseline hazard estimator) tracked CoxPH closely and outperformed RSF and GBSA across censoring levels.
Generated synthetic survival data with controlled censoring to stress test model behavior under different conditions.

Python · PyTorch · scikit-survival · survival analysis

GitHub

Deep learning for remaining useful life

Battery RUL CNN

5-block 2-D CNN (PyTorch) regressing lithium-ion battery remaining useful life from interpolated cycle discharge curves.

R² 0.76 on held out test batteries — RMSE 148 cycles, MAE 87 cycles.
Converted raw cycle discharge curves into model ready 3 channel image style inputs.

PyTorch · CNNs · RUL prediction · time-series preprocessing

GitHub

Earlier exploration — EPFL & Yeshiva University

MATLAB Predictive Maintenance Workflows in Python

Ported MATLAB Predictive Maintenance Toolbox workflows to Python — reliability modeling and wind turbine bearing degradation. The battery CNN above grew directly out of this work.

Python · reliability modeling · MATLAB-to-Python translation

GitHub

Technical focus: applied ML systems, industrial AI, RAG/LLM extraction, physics simulation, reliability modeling, uncertainty quantification, and the systems work to run it in production — streaming ingest, multi-tenancy, and deployment.