David Papp
AI solution developer
AVAIL // Full-time
| LOCATION | : | Amsterdam · Rotterdam, NL · remote |
| FOCUS | : | ai agents · rag · solution dev |
| CURRENTLY | : | building AI solutions at WEBINFORM & studying Econometrics and Data Science at VU Amsterdam |
| AVAILABILITY | : | Full-time |
| WORK AUTH | : | NL / EU Work Authorization (No sponsorship required) |
| WORK MODE | : | Remote / Hybrid / On-site |
| LANGUAGES | : | English (Fluent), Hungarian (Native) |
| CONTACT | : | contact@davidpapp.dev |
4 active nodes found — click row to expand case study
Interactive portfolio demos — illustrative by design, hosted as independent static bundles.
An interactive chat where the site interviews David — deterministic answers over verified facts only.
Open Self-Interview in a full pageDescribe a task and watch an illustrative automation workflow assemble, with assumption-based effort estimates.
Open Task-to-Flow in a full pageA short quiz that maps your context to an industry lens and suggests which work fits — illustrative, not a scored assessment.
Open Fit-Finder in a full pageI build AI solutions professionally at WEBINFORM — production LLM features, APIs, and automation for web applications and ERP-integrated systems — and, on my own time, local-first context infrastructure, agent orchestration tooling, and the integrated AgentSec security suite. I began university studies in 2024 and changed to Econometrics and Data Science at VU Amsterdam in September 2026; graduation is expected in 2028. I'm looking for full-time AI engineering, AI solutions, integration, automation, or agent-infrastructure roles.
- > Build internal AI tools and production LLM functionality for web applications and ERP-integrated systems: APIs, automation, backend logic, and user-facing AI interfaces.
- > Recovered an inherited AI-first service by reverse-engineering undocumented integrations, repairing architecture and security issues, and reducing LLM API costs. Addressed a prompt/context-injection vulnerability using deterministic validation and human review.
- > Delivered client websites and webshops, internal systems, user-facing AI functionality, and ERP/AI integrations, with direct involvement in discovery, solution design, client coordination, and pricing.
- > Coursework includes machine learning, computational linguistics, multi-agent systems, and the mathematics underneath them.
- > Taught Python and Unity using pre-made curricula.
- > Taught JavaScript through a visual block-based tool.
- > Drafted a first operating procedure for sorting, tracking, and handoff on a warehouse line, from engineer walkthroughs.
- > Helped trace warehouse-line process gaps and logged findings for the engineers. Supported sorting and distribution handoffs on the warehouse line.
WORK = used in professional client delivery · LAB = personal open-source projects · UNI = university coursework
- ● LLM API IntegrationWORK
- ● Prompt & Context EngineeringWORK
- ● Agent Hooks & Policy ToolingLAB
- ● RAG PrototypesLAB
- ● PythonWORK
- ● SQL / PostgresWORK
- ● Machine Learning FoundationsUNI
- ● Vector & Semantic SearchLAB
- ● Next.js / TypeScriptWORK
- ● APIs & Backend LogicWORK
- ● ERP IntegrationsWORK
- ● RustLAB
Engineering notes from building the projects above — each links to the repository where the pattern is implemented
Fail-closed decision hooks for Claude Code and Codex: safe-command fast paths, chained-command bypass protection, and observe/prompt/enforce modes. Implemented and tested in the repo.
Allowed-roots path validation, destructive-command deny patterns, and regex credential redaction of subprocess output — the policy and redaction modules carry the unit tests.
Accepts a raw MCP tools/list response and returns capability risks, severity, and missing security controls using transparent, explainable rules — it never executes or probes the listed tools.
Stores declared agent metadata and computes a 0-100 risk score from permissions, data access, and tool descriptions — readiness labels and human-readable factors, not runtime enforcement.
Classifies an action's risk deterministically, queues it for a human decision, and records approve/reject with reasons — the calling system waits for and enforces the recorded decision.
A small rule set flags instruction overrides, jailbreak markers, and exfiltration intent with evidence excerpts and remediation guidance — an auditable first-pass filter, honestly labelled heuristic.
I'm open to full-time engineering and AI roles. Based in Rotterdam, NL — remote / hybrid. No sponsorship required. Email is fastest.
Session active. Grounded in David's reviewed portfolio knowledge base. Ask me about his engineering work.