AI/ML Engineer · open to roles
Taya Yakovenko
I build AI systems that pair models with human judgment, end to end, for work in high-stakes environments.
About
My work runs across domains that don’t usually share a resume: fine-tuning LLMs for credit-risk scoring, catching anomalies in live network traffic, finding regions of interest in microscopy images, building LLM apps people actually use. What ties them together is the kind of problem I’m drawn to. Ones where a wrong answer is expensive, and where the model on its own isn’t enough.
In those settings I keep landing on the same shape of solution. The model does the heavy, repetitive work; a person keeps control of the call that matters. That’s usually not a compromise, it’s the design that performs best, so I build the whole loop around it: the data pipeline, the evaluation that proves it works, and the point where a human makes the final decision.
That instinct comes from a double background in math and computer science. The math is why I trust statistical detectors, uncertainty, and honest evaluation before reaching for a neural net; the CS is why the result ships as a real system instead of a notebook. New domain or new stack, the approach stays the same.
Experience
-
Risk LLM Fine-Tuning (NDA) · PayPal
Remote
Credit-risk assessment with fine-tuned LLMs under strict confidentiality.
- Built an end-to-end LLM fine-tuning workflow (data prep + prompt design) using LoRA/QLoRA for credit-risk assessment.
- Benchmarked fine-tuned LLMs against classical baselines (Random Forest, SVM, XGBoost) on F1, KS, and MCC with multi-seed stability checks.
PyTorchTransformersLoRA/QLoRAscikit-learnAzure AI FoundryHPC -
ML Researcher (Computer Vision) · UT Health San Antonio — Klykov Lab
Remote · part-time
Operator-in-the-loop ROI detection for cryo-CLEM microscopy targeting.
- Building a zero-shot, ML-assisted workflow to identify regions of interest in 3D cryo-fluorescence brain-slice volumes, cutting the manual expert annotation that bottlenecks astrocyte-related cryo-EM targeting (no experiment-specific training data).
- Designed a four-level hierarchical pipeline: zero-shot segmentation (Cellpose-SAM) → per-ROI semantic scoring with a biomedical VLM (BioMedCLIP) → rule-based spatial plausibility → a napari curation viewer that logs every accept/reject to schema-versioned, machine-readable provenance.
- Set up the lab’s remote compute access (SSH into a pool of lab servers over a Hamachi VPN) so the pipeline runs on lab GPUs from anywhere.
PythonCellpose-SAMBioMedCLIPnapariscikit-imageDocker -
Data Analyst (ML Performance & Reporting) · Empower Analytics
Worcester, U.S. · part-time
Pricing-optimization analytics for transportation & logistics clients.
- Stabilized a live performance dashboard for an ML pricing tool by migrating reporting logic from Google Sheets to Python (Spark/Pandas) and adding data validation for inconsistent client inputs.
- Analyzed edge cases and data-quality issues to improve feature reliability and reduce dashboard breakages during live updates.
- Produced weekly model-performance reports with visualizations and t-tests for client updates.
PythonPySparkPandasSQLMatplotlibD3.js -
Software Developer Intern · Hanover Insurance Group
Worcester, U.S.
Property & casualty insurer — claims data automation.
- Built a Python/Pandas automation to parse Excel-based claim data into an internal-system-ready format, cutting manual entry time (~15× faster vs. manual processing).
- Supported data preparation for a Power BI reporting workflow.
- Worked within an Agile sub-team of 5, delivering increments and integrating scripts into the internal flow.
PythonPandasExcelPower BI
Skills
Machine Learning
- LLM fine-tuning (LoRA/QLoRA)
- Model evaluation
- PyTorch
- scikit-learn
- Computer vision
- Anomaly detection
- Statistical modeling
AI-Assisted Development
- Claude Code
- Agent orchestration
- Context engineering
- Skills & MCP setup
- Multi-agent workflows
- LLM APIs (Claude, Gemini, Qwen)
Engineering & Infrastructure
- Python
- Go
- Linux
- Docker
- Kubernetes
- Git
- CI/CD
- REST APIs
Data & Cloud
- SQL
- PySpark
- Pandas
- MongoDB
- Azure ML
- HuggingFace
Projects
Selected work
Each one started as a real problem and ended as something that runs.
Flagship ROI-ID · Astrocyte
Operator-in-the-loop ROI detection for cryo-CLEM (correlative cryo-light/electron microscopy). A hierarchical, zero-shot pipeline that finds targets in cryo-fluorescence volumes to aim the cryo-EM that follows.
MIRA
Real-time BGP routing-anomaly detection: a dual statistical detector on a concurrent Go streaming pipeline, surfaced through a REST API and a live dashboard.
Live demo BlogAI
A deployed LLM drafting app that turns an article plus your notes into a blog and LinkedIn draft. Live on HuggingFace Spaces, with a human in the loop.
Eval study BlogAI Evals
An evaluation of whether the BlogAI pipeline keeps what matters when notes become a post: the user’s argument and their voice. The question underneath is whether the cheap Qwen→Haiku pipeline holds up against Haiku on its own.
Pricing Model
A human-in-the-loop pricing model for a logistics firm: it suggests freight quotes from historical data, learns from every approve / reject / correction, and retrains in batches.
In progress objectDimensions
A computer-vision pipeline aiming at clinical-grade 3D body measurement from ordinary phone video, fusing IMU physics, reference geometry, and a statistical body model.
Live demo Night-to-Day
A generative U-Net for low-light image enhancement, with a live demo, dataset, and multiple model versions published on HuggingFace.
Live demo The Automation Paradox
A data study on AI’s real impact on the US job market: regression, clustering, an ETL refresh pipeline, and a live interactive dashboard.
Coursework & foundations (4)
- VAE Architecture Comparison Comparative study of variational-autoencoder architectures.
- Weather Classification (Semi-Supervised) K-Means pseudo-labeling and recursive SVM confidence thresholding on spatial pixel statistics — no neural nets.
- TPC-H Query Efficiency Query-optimization study over the TPC-H benchmark.
- Data Visualization & Perception Perception-driven visualization study.
Contact
Let’s talk.
Open to ML engineering roles. The fastest way to reach me is email.