Projects

Selected work, from production AI platforms to systems programming

Flagship Project

Vetdux

AI-driven anamnesis platform for veterinary care
Lead AI & Software Architect at Norclin, Nov 2025 - Present

Contracted to design and build Vetdux, an SLU Holding-funded platform where every veterinary consultation starts from a finished anamnesis: structured intake plus an AI-led interview, with an always-on safety guardian watching for emergencies. I scoped the product into seven independently deliverable modules, produced the business case, and led architecture and delivery end-to-end: product, AI engine, backend, two frontends, cloud, security, and compliance. Along the way I mentored a junior developer, an AI researcher with whom I collaborated on the evaluation methodology.

Vetdux AI-led clinical interview screen in perspective view

AI Engine

  • 14-node LangGraph interview engine with parallel and speculative LLM execution, roughly halving per-turn latency.
  • Structured LLM outputs throughout: every model call returns validated Pydantic objects, and raw model text is never parsed.
  • Hand-authored clinical decision trees combined with generative interviewing: deterministic at zero LLM cost, generative when no tree matches.
  • Embedding-based semantic routing matches complaints to decision trees across languages via cosine similarity.
  • Always-on safety guardian beside the interview: emergency detection, clinical-signal classification, and bailout handling.
  • Multilingual AI orchestration with English reasoning and localized output; adding a language is a configuration change.

Architecture & Product

  • Strictly layered async Python backend: FastAPI, Strawberry GraphQL, Protocol boundaries, scoped dependency injection.
  • The interview is one server-side state machine with a frozen set of legal transitions, so illegal states crash instead of corrupting data.
  • Two fully server-driven Expo/React Native web apps: a patient app and a portal that runs embedded in Provet Cloud.
  • Dual-mode integration architecture (standalone and Provet Cloud PIMS) behind a strategy interface, including booking flows and journal write-back.
  • Perceived performance engineered for multi-second LLM latency with staggered micro-animation waterfalls, honoring reduced-motion preferences.

Cloud & Security

  • Multi-account AWS plus GCP Vertex AI built entirely as code with Terraform: pausable stacks and 13+ reusable modules.
  • Keyless multi-cloud identity with GitHub OIDC and AWS-to-GCP Workload Identity Federation, with no long-lived credentials.
  • Transport secured end-to-end: Cloudflare full-strict TLS, mTLS origin pulls, backend reachable only via loopback.
  • CI/CD as a dozen GitHub Actions workflows: secret scanning, SAST, IaC scanning, dependency audits, tests, and gated deploys.
  • Authentication in depth: argon2id hashing, pinned-algorithm JWTs, constant-time HMAC verification, strict multi-tenant isolation, and layered prompt-injection defenses.

Compliance & Quality

  • GDPR and EU AI Act compliance implemented as working code: seven-tier retention with explicit legal bases, automated data-subject-rights flows, versioned HMAC-signed consent.
  • Multi-gate anonymization pipeline (k-anonymity, l-diversity, and an automated motivated-intruder LLM probe) that gates all deletion.
  • LLM evaluation harness with 2,689 synthetic clinical cases across 12 species groups, including 106 adversarial cases spanning 13 attack types.
  • Load generation with token-bucket rate control and resumable runs, reporting p50/p95 latency toward a 1000+ concurrent-session target.
  • 23,000+ lines of design docs, 15+ runbooks, and decision records; production-ready across three isolated cloud environments.
Vetdux safety check screen asking whether the visit is an emergency
Every session starts with a safety check for emergencies.
Vetdux appointment booking screen with calendar and available time slots
Integrated appointment booking against the clinic's calendar.
Vetdux booking confirmation screen
The vet starts the confirmed visit from a finished anamnesis.

Technologies

PythonFastAPIGraphQLLangGraphPydanticGemini / Vertex AIPostgreSQLReact Native / ExpoTypeScriptTerraformAWSGCPGitHub ActionsCloudflareGDPREU AI Act

More Projects

AI Agentic Problem Solver

  • Engineered a hierarchical multi-agent system (Architect, Planner, Executor agents) for complex problem-solving through systematic decomposition and specialized task handling.
  • Leveraged LLMs within each specialized agent for high-level design, detailed planning, and code implementation.
  • Built on a protocol-based, modular architecture with distinct components for configuration, agent management, LLM provider integration, and messaging.
  • Incorporated robust error handling, retry mechanisms, timeout management, and fallback delegation paths for reliable, autonomous task execution.
  • Implemented agent-driven delegation where agents assess task complexity to dynamically route work across the hierarchy.
PythonLLMsLangChainLangGraphasyncioCustom Agent Framework
View on GitHub

Cadenzra - AI Music Composition Pipeline

  • Built a notation-first AI music composition pipeline where every run emits five mutually consistent artifacts: engraved PDF score, LilyPond source, score MIDI, interpreted performance MIDI, and rendered audio.
  • Compiles free-text style prompts into concrete, machine-checked musical constraints: every bar must fill its declared meter and every note must stay within its voice's declared range, failing loudly instead of emitting invalid output.
  • Supports odd and additive meters, long-form structural planning with thematic returns, and lyric coordination across multiple voices.
  • Modular LLM provider layer with prompt-quality evals run against live models.
PythonAI/LLMsGeminiLilyPondFluidSynthMIDIpytestDocker
View on GitHub

WaveForge - GPU-Accelerated Terrain Generation

  • Developed a high-performance implementation of the Wave Function Collapse (WFC) algorithm for 3D terrain generation.
  • Engineered for GPU acceleration using WGPU compute shaders, significantly improving generation speed.
  • Designed a modular architecture with distinct layers for API, coordination, implementation, and hardware abstraction.
  • Implemented a flexible rule definition system for controlling terrain characteristics, with visualization of the generation process.
RustWGPUWGSLCompute Shaders
View on GitHub

Quirefold - Structurally Clean Markdown-to-PDF Pipeline

  • Built a Markdown-to-PDF pipeline that guarantees structural integrity: a recursive packing algorithm measures every node's rendered height by test-compiling and groups content into page-sized unbreakable units.
  • Eliminates mid-paragraph, mid-sentence, and mid-list-item page splits, with instrumented quality metrics verifying every run against an explicit whitespace budget.
  • Integrates pandoc, Typst, and pre-rendered Mermaid diagrams into a containerized one-command pipeline.
PythonTypstpandocMermaidDockerAlgorithm Design
View on GitHub

Tallynix - Swedish Invoice Generator CLI

  • Developed a Rust CLI that renders Swedish-compliant invoices (organization number, VAT rates, payment terms) to PDF via the Typst typesetting system.
  • Produces dual output: a human-readable PDF and a machine-readable JSON record of all computed amounts.
  • Handles line items with display-only sub-items kept strictly separate from computed totals, ensuring correct VAT and monetary arithmetic.
RustTypstJSONPDF Generation
View on GitHub

Wedding Website - saraoliver.se

  • Designed, built, and launched a full-featured wedding website with RSVP and accommodation booking, including per-room capacity limits and automatic per-person cost splitting.
  • Implemented two-tier access control: a site-wide invite gate with HMAC-signed cookies and constant-time verification, plus a Discord-OAuth admin portal for managing submissions.
  • Engineered the site GDPR-by-design: explicit Article 9 consent for dietary data, strictly necessary cookies only (no cookie banner needed), and documented retention and data-subject-rights procedures.
  • Built deep-linkable single-page navigation with scroll-snap sections and IntersectionObserver-driven URL synchronization.
Next.jsTypeScriptReactPrismaTursoBetter AuthZodTailwind CSSVercelGDPR
Visit Website

Professional Portfolio Website - linneasvensson.com

  • Designed and developed a personal CV and portfolio website for a Quality Assurance professional.
  • Focused on a clean editorial design, responsive layout, and content the owner can easily keep current.
Next.jsReactTypeScriptTailwind CSS
Visit Website

Aeriasong

  • Led development as the lead developer for a 2+ year game project.
  • Architected and implemented core game systems, gaining broad experience across the full development lifecycle from concept to implementation.
GodotGDScriptC++Game Development
Visit Website

Parabel Game Studio Website

  • Developed and launched the official website for Parabel Game Studio, showcasing company information, projects, and team details.
  • Includes a playable WASM game demo directly in the browser.
  • Used for distributing beta keys for the game, with cookies tracking who has already claimed a key.
Next.jsReactTailwind CSSWASMJavaScript
Visit Website

Campaigen - Marketing Campaign Management CLI

  • Developed a .NET 8 CLI application for managing marketing campaign data, focusing on spend tracking and influencer details.
  • Engineered using Clean Architecture principles combined with Vertical Slicing, with distinct Domain, Application, Infrastructure, and Presentation layers.
  • Implemented data persistence using Entity Framework Core with an SQLite database.
C#.NET 8Entity Framework CoreSQLiteClean Architecture
View on GitHub

Spantry - Smart Pantry CLI

  • Developed a command-line application for pantry inventory management: items with name, quantity, location (fridge, pantry, freezer), and expiration dates.
  • Built with Java 17, Gradle, and Picocli, with JUnit 5 tests and Jakarta Bean Validation.
Java 17GradlePicocliJUnit 5
View on GitHub

Personal Portfolio Website

  • Designed and developed this personal portfolio website to showcase projects and skills.
  • Focused on a clean, responsive design and user experience.
Next.jsReactTypeScriptTailwind CSS
Visit Website