ML engineer · Independent consultant · Cluj, Romania

I build ML systems that find the signal in noisy data.

Six years shipping production ML in fintech, mapping, defence and telecoms: time series, anomaly detection, computer vision and LLM/RAG. Trained as a physicist (MPhys, CERN research). Currently taking on consulting work.

Get in touch LinkedIn GitHub

PreviouslyBanca Transilvania · Grab · Rinicom · AccessPay · LHCb at CERN · University of Manchester

Selected work

Recent projects.

Fintech · On-device computer vision · Nov 2025 – Apr 2026

Android document scanner for a fintech platform

Replaced the brittle OpenCV detector in an invoice-processing app's Android scanner with a U-Net/MobileNetV2 segmentation model that runs on budget phones like the Samsung A12. In production since April 2026.

99%: pixel accuracy on unseen images; the old detector managed ~70%
30–50 ms: preprocessing on a low-end phone, down from 700–1000 ms
600 → 300: lines of code, heuristics replaced by one pipeline

Telecoms · Data

Maintenance deconfliction

Stops planned maintenance from hitting diverse network circuits at the same time. Dataiku and Snowflake, over inventory data too big for memory.

Compliance · RAG evaluation

Auditing a regulatory RAG system

Showed the vendor's metrics were flawed and rebuilt the evaluation on real-world data. The findings changed the client's strategy.

Healthcare · Computer vision

Image similarity for a blind study

Grounding DINO detection and CLIP embeddings in ChromaDB, with face anonymisation and a Plotly Dash front end for blind evaluation.

Banking · RAG

Chunking for long procedures

Bottom-up hierarchical chunking for a bank's long internal procedures. Retrieval recall doubled.

Internal docs · RAG

Internal knowledge assistant

RAG over a company's internal documentation for employees, extended with graph-database lookups.

Engineering · AI-assisted development

Multi-agent coding workflow

A Claude Code setup with analyst, codebase-mapping, architect, implementer and verifier agents. Open decisions come up, with suggested answers, before any code is written.

Writing

Analyses I run for myself, charts and method in the open.

VWCE or Value + Quality? What 56 years of crises say
Nov 2026
A factor portfolio does not protect you in a crash. On paper it wins over the long run; the actual funds have trailed the market.

Working together

I help teams take ML models into production: forecasting and anomaly detection on time series, computer vision on mobile devices, and RAG systems evaluated on real data.

hello@vladgriguta.com