The work ofIoan-Vlad Pantiru

ABOUT

Ioan-Vlad Pantiru

AI engineer shipping production LLM systems end to end: LangGraph agents, RAG and Azure OpenAI, from data pipelines and verification gates to full-stack apps and CI/CD. Former technical co-founder of an edge-AI startup.

• Cluj-Napoca based in
• EU remote open to

EDUCATION

From AI to security

BSc in Artificial Intelligence at Babeș-Bolyai University, 2023–2026. Now on the EIT Digital double-degree MSc in Cybersecurity: year one at UBB, year two at the University of Twente.

• 2028 MSc expected
• C1 English, CSE 199

2024–25 · M100-CITIES HACKATHON

WalkSafe

Pedestrian routing that weighs safety and air quality, not just distance. React Native, Django, PostgreSQL, OpenAI, scikit-learn and osmnx.

• Special Mention

2025–26 · TECHNICAL CO-FOUNDER

Horizon HUD

Fully offline perception on a Raspberry Pi 4: YOLOv8s trained on 70K BDD100K images and quantized to INT8, SORT tracking, ego-motion compensation and a risk engine ranking hazards on a helmet display.

• 6 FPS on-device
• 70K training images
• 306 tests on the Pi

LIVE · HORIZON RISK ENGINE

Perception feed

Animated perception feed: a road in perspective with cars and motorbikes approaching, each tracked with a labelled bounding box. Vehicles in the rider's lane past mid-distance are flagged as risk. Model yolov8s-int8 with SORT tracking and ego-motion, running at 6 FPS on a Raspberry Pi 4.

THE HARD CALL

Knowing when to pause

Finalist at Innovation Labs 2025 and Stage Two in Berlin, and the basis of my bachelor thesis. Then a market analysis showed rising hardware costs pushed the price past what riders would pay, so we paused it.

COMPUTACENTER · LEAD DEVELOPER

AI IT asset platform

Turns 20K+ device client inventories into specs, refresh plans and ROI reports. Shipped to production in about two months and used by around fifteen consultants.

• ~$600K savings per client
• 2d → 1h research time
• ~80% dates resolved

COMPUTACENTER · LANGGRAPH

The pipeline, live

Animated pipeline: ingest, map, research, gate and ship. A vendor API and Bing feed the research step. Records that fail the gate branch off and return TBC. Unverified fields come back as TBC instead of a confident guess.

THE QUALITY BAR

Built to be trusted

LLM task routing and telemetry across 13 AI modules, Celery/Redis batch enrichment, Entra ID SSO, around 1,300 automated tests, and CI moved to self-hosted ARM64 runners to cut cost.

• 13 AI modules
• ~1,300 tests

AZURE AI FOUNDRY

More systems shipped

A RAG recruitment assistant on pgvector and Redis, integrated with SharePoint and SuccessFactors. A ServiceNow AI Copilot. A multi-agent device-data enrichment system.

• 3 Foundry solutions

FREELANCE · SOLE DEVELOPER

MitMed clinic platform

Client portal and back office for mitmed.ro. The clinic went fully paperless: booking with Google Calendar sync, WhatsApp reminders, GDPR consent management and PayU billing.

• ~3 wk to live
• 60–70 clients per month
• 130+ tests, CI/CD

FREELANCE · IN BUILD

Construction platform

An end-to-end system for a mid-size builder: stock, crews, projects and cost control. LLM document extraction, and bid drafts where every figure cites a company record.

• e-Factura integration
• REGES integration

1ST PLACE · GREENPOW · TALLINN 2026

HubShift

With team Watt Matters: a GPU pod that runs AI batch jobs on the idle grid capacity of EV fast-charging hubs and yields to cars within seconds. Pitch, business model, live dashboard and pilot plan.

• 1st place
• ~94% idle capacity
• 6 teammates

TOOLKIT

What I build with

LangGraph · LangChain · Azure OpenAI · AI Foundry · RAG · pgvector · YOLOv8 · TFLite INT8 · OpenCV · Python · TypeScript · FastAPI · Next.js · React Native · PostgreSQL · Celery · Redis · Docker · GitHub Actions · Entra ID

THE LAST ROOM IS YOURS

Let's build something that ships.

ioanvladpantiru@gmail.com · LinkedIn · GitHub · open to EU remote