ENGINEERING, WITH INTENT.CLUJ-NAPOCA, RO 46.77° N / 23.59° E

Hi, I’m David.

I build AI.
And everything
around it.

AI/ML Engineer & Full Stack Developer

From a retrieval problem to a working application. I connect models, evaluation, backend systems, and cloud delivery.

MY AI DEVELOPMENT FLOW Explore

01

Start with data you can trust.

Garbage in, garbage out. I collect relevant inputs, clean and validate them, and prepare a consistent foundation for the model.

Currently AI Engineer @ RebelDot / Nov 2024–present

A little about me

MY EVERYDAY BUILDING BLOCKS

Tools with a purpose.

Pick a tool. See where it fits.

Python

The connective tissue across my AI systems, APIs, and evaluation pipelines.

Used across RAG, call quality assurance, and platform engineering.

Some of my results.

The numbers behind the work. Explore each story.

EMAIL KNOWLEDGE PIPELINE45% → 88%Reported correct drafting CALL QUALITY EVALUATION952 judgments8 criteria across 119 calls PLATFORM OPTIMIZATION25s → <2sQueries over 2.5 million rows

THE WORK / 2024–PRESENT

Engineering choices.
Visible outcomes.

AI Engineer at RebelDot.
Real systems, personal ownership, and the evidence behind the work.

RebelDot

AI Engineer · Cluj-Napoca, Romania

Nov 2024–present

01AUTOMOTIVE MARKETPLACEAug 2026–present

RAG-based Email
Response Automation

From past replies to reusable knowledge.

I independently architected and implemented an 11-category RAG assistant. It retrieves relevant knowledge and generates source-cited email drafts for human review.

PythonDSPy / GEPAAWS BedrockPostgreSQL / pgvector

AGENT-REPLY KNOWLEDGE PIPELINE

Correct drafting

45% 88%

Top-5 retrieval

70% 91%

Reported measurements for the integrated pipeline.

What I built

Retrieval with the right context. Routed requests across 11 categories, combined Cohere embeddings with PostgreSQL/pgvector search, and used Haiku reranking to prioritize evidence before drafting.
A reusable knowledge pipeline. Turned historical agent replies into source-grounded knowledge. The integrated pipeline raised reported correct drafting from 45% to 88% and top-5 retrieval from 70% to 91%.
Preparation before indexing. Added GLiNER/Presidio PII masking, MPNet semantic clustering and DeBERTa NLI grounding checks to limit personal-data exposure and keep answers traceable to their sources.
Repeatable prompt optimization. Replaced ad hoc tuning with DSPy/GEPA against defined classification and safety metrics, making prompt selection repeatable and quality changes measurable.

02INSURANCE2026

LLM-based Call
Quality Assurance

Assessments you can actually inspect.

I built weighted, rubric-driven LLM scoring with evidence for each criterion. RapidFuzz and DeBERTa NLI flag fabricated citations and unsupported reasoning without extra LLM verification calls.

PythonLangChainDSPyHuggingFace Transformers

952judgments

8 criteria · 119 calls

0102030405060708

Evaluated with macro-F1 and Cohen’s kappa to track agreement and regressions.

What I built

Scoring with inspectable evidence. Turned transcripts into weighted, rubric-driven assessments with citations for each criterion, making quality reviews consistent and scores auditable.
Verification without extra LLM calls. Combined RapidFuzz citation checks with DeBERTa NLI to flag fabricated citations and unsupported reasoning. Local NLI checks benchmarked at 108–151 ms per criterion; these are advisory checks, not a factuality guarantee or total application latency.
Optimization against human agreement. Used DSPy/GEPA with macro-F1 and Cohen’s kappa across eight criteria and 119 calls (952 judgments) to measure agreement and identify quality regressions.

03AI APPLICATIONS & PLATFORM ENGINEERINGNov 2024–Dec 2025

Strong foundations.
Better applications.

Across insurance, automotive, and enterprise work, I delivered cloud AI applications and improved the systems they depend on.

FastAPIStreamlitAWS / AzureTerraform

25s → <2s

Queries across 2.5 million rows with PostgreSQL, batched APIs, and Redis.

3 → 40+

Filtering parameters, alongside 19 self-scoped optimization tickets under a Q4 deadline.

What I built

Faster access to 2.5 million rows. Delivered 19 self-scoped optimization tickets, using PostgreSQL tuning, materialized views, batched API endpoints and Redis caching to bring queries from 25 seconds to under 2 seconds.
Filtering that could grow. Refactored API and UI architecture with Builder and component-based patterns, expanding from three to 40+ cascading filters under a Q4 deadline, with tests and incremental retirement of older services.
More efficient application execution. Profiled memory use, fixed leaks and improved cache management. Added LLM caching and cross-region AWS Bedrock load balancing to reduce repeated initialization and throttling pressure.
Applications connected end to end. Delivered FastAPI and Streamlit applications with Microsoft Entra ID/JWT authentication, processing feedback, Terraform infrastructure and Azure Pipelines delivery.

MORE FROM THE SAME PERIOD / NOV 2024–DEC 2025

AI, put to work.

More of the applications behind that foundation.

Multimodal document OCR

OutcomeTurned PDFs and images into structured Markdown and Excel downloads through an authenticated API and an upload interface.

Extraction pipeline: Built pdf2image/Poppler conversion, multithreaded batched AWS Bedrock inference, result synthesis and output generation.
Extensible extraction: Designed modular prompts so new document use cases could be added without changing the core inference pipeline.
Usable delivery: Built FastAPI endpoints, JWT/Microsoft Entra SSO, presigned S3 downloads and a Streamlit interface with uploads and processing feedback.
Repeatable releases: Provisioned AWS infrastructure with Terraform and configured Azure Pipelines CI/CD with temporary environments.

PythonAWS Bedrockpdf2image / PopplerFastAPIStreamlitAWS S3TerraformAzure Pipelines

AI-assisted vehicle listings

OutcomeSecured access to a vehicle-description assistant, reduced repeated authentication and model setup, and made vehicle inputs more tolerant of typing errors.

Protected access: Integrated frontend SSO and backend token validation for the FastAPI service.
Fewer external requests: Implemented JWT and long-lived token caching to reduce repeated authentication calls and memory use.
Reusable model setup: Cached LLM chains and model instances to avoid rebuilding them for each request and reduce application load.
More forgiving inputs: Built typo-tolerant matching for vehicle make, model and year inputs.

PythonFastAPIStreamlitAWS BedrockSSO / JWTLLM cachingPodmanAzure Pipelines

Workforce & sales assistant

OutcomeHelped take role-specific workforce and sales assistants from concept to deployment in three months.

Automated infrastructure: Provisioned Azure services with Terraform and automated releases through Azure DevOps/Azure Pipelines.
Container delivery: Packaged applications with Docker, published images to Azure Container Registry and deployed them to Azure App Service.
Connected services: Managed environment configuration and Azure service connections for the application’s cloud services.
Role-specific access to information: Configured OpenWebUI assistants for finance and sales teams to work with workforce availability and sales information.

PythonOpenWebUIDockerTerraformAzure App ServiceAzure Container RegistryAzure PipelinesAzure Key Vault

THE TOOLS BEHIND THE WORKExplore my toolkit

TOOLKIT / INTENT BEFORE TOOLS

The right tool.
For the real problem.

My strongest work happens at the intersection of AI, data, backend engineering, and cloud delivery.

01 / AI ENGINEERING

Quality you can measure.

Grounded answers, repeatable optimization, and evaluation that makes regressions visible.

PythonRAGDSPy / GEPALangChain / LangGraphAWS BedrockClaudeStructured outputsEvaluation

02 / RETRIEVAL & NLP

Evidence, in context.

Relevant retrieval, sensitive-data masking, reranking, and traceable sources.

PostgreSQL / pgvectorHuggingFace TransformersCoherePyTorchDeBERTa NLIMPNetGLiNER / Presidio

03 / BACKEND & DATA

From model to application.

Typed APIs, reliable data access, caching, and interfaces people can use.

FastAPIPydanticSQLAlchemySQLRedisStreamlitpandas / NumPy

04 / CLOUD & DELIVERY

Built to keep moving.

Infrastructure, identity, testing, and delivery practices that support maintainable systems.

AWS / AzureTerraformDockerAzure PipelinesGitpytestRuff / mypy

HOW I WORK

Good engineering is a team sport.

I established shared coding-agent rules and pre-commit checks for style, tests, and commits. Architecture and contribution guides I wrote were adopted by the team, making technical collaboration clearer.

TOOLS, APPLIEDSee them in the work

ABOUT / THE HUMAN PART

Curious by nature.
Engineer by practice.

Cluj-Napoca, Romania.
Romanian & English.
Usually thinking about the next problem—or the next run.

01 / MY APPROACH

I like owning
the whole puzzle.

From the first unanswered question to the working application. I enjoy shaping the architecture, solving the awkward details, and communicating the tradeoffs.

My work combines initiative, analytical problem-solving, and technical communication with a practical focus on delivery under real deadlines.

See how that translates into work

OFF THE KEYBOARD

Running. Triathlon.
Steady progress.

I’m registered for IM 70.3 Mamaia 2027. A future goal I’m working toward, one session at a time.

ON THE HORIZON / 2027

LEARNING, PUT TO WORK

Credentials & foundations.

01

CERTIFICATION

Claude Certified Architect – Foundations

Claude API/Agent SDK, agentic orchestration, tools and MCP integration, structured outputs, context management, and reliability.

20 June 2026–June 2027

02

CERTIFICATION

Microsoft Certified: Azure AI Engineer Associate (AI-102)

Azure AI design and deployment, NLP, computer vision, document extraction, search, and responsible AI.

Earned 2024 · Renewal status unconfirmed

03

EDUCATION

Technical University of Cluj-Napoca

Bachelor’s Degree in Electronics, Telecommunications and Information Technology.

2020–2024

START A CONVERSATIONLet’s connect

CONTACT / GOOD THINGS START HERE

Something
on your mind?

AI engineering, a useful product, a hard problem.
Let’s talk about what we could build together.

STRAIGHT TO MY INBOX

farcas.david@yahoo.com

+40 720 806 333

I build AI.And everythingaround it.

Start with data you can trust.

Tools with a purpose.

Python

Some of my results.

Engineering choices.Visible outcomes.

RebelDot

RAG-based EmailResponse Automation

What I built

LLM-based CallQuality Assurance

What I built

Strong foundations.Better applications.

What I built

AI, put to work.

Multimodal document OCR

AI-assisted vehicle listings

Workforce & sales assistant

The right tool.For the real problem.

Quality you can measure.

Evidence, in context.

From model to application.

Built to keep moving.

Good engineering is a team sport.

Curious by nature.Engineer by practice.

I like owningthe whole puzzle.

Running. Triathlon.Steady progress.

Credentials & foundations.

Claude Certified Architect – Foundations

Microsoft Certified: Azure AI Engineer Associate (AI-102)

Technical University of Cluj-Napoca

Somethingon your mind?

I build AI.
And everything
around it.

Engineering choices.
Visible outcomes.

RAG-based Email
Response Automation

LLM-based Call
Quality Assurance

Strong foundations.
Better applications.

The right tool.
For the real problem.

Curious by nature.
Engineer by practice.

I like owning
the whole puzzle.

Running. Triathlon.
Steady progress.

Something
on your mind?