All Work
Archive.

An Austrian company's institutional knowledge was buried across thousands of internal documents, wikis, and PDFs — staff couldn't find answers without interrupting a senior colleague.

Over a four-month remote engagement with the Austrian team, I built a retrieval-augmented knowledge assistant: hybrid search over embedded internal documents, a reranking layer for precision, and grounded LLM answers with source citations so every response could be verified and trusted.

Two Norwegian startups needed to move from idea to shipped AI features fast, without large teams to lean on.

Across nine months I integrated with two different teams — fast enough that it felt like I'd always been there — and shipped an LLM-powered chat assistant, a RAG document-search tool, and generative content pipelines. Tight cycles: build, test, ship in days, iterating with real users before hardening anything for production.

Legal and procurement teams lose hours manually reading contracts, extracting key terms, and checking obligations across long PDF documents.

Co-built a three-tier platform: a Next.js client (NextAuth, Prisma), an Express core API, and a Flask AI service running Mistral OCR and Mistral-Large through a LangChain pipeline to turn raw contract PDFs into structured, queryable data.

RAG systems are only as good as their inputs, and real-world PDFs — scans, tables, multi-column layouts — break naive text extraction.

Built a production document-extraction pipeline: EasyOCR workers, a model for table and figure detection, rule-based validators, and orchestration that feeds clean, structured chunks into a Flowise RAG stack. Dockerized and deployed on Railway.

Responding to public tenders means reading hundreds of pages of requirements, scoring candidate CVs against rigid criteria, and drafting a compliant technical memo — days of expert work per bid.

Designed a FastAPI service that OCRs tender documents, extracts the scoring grid, scores candidate CVs against each required role, and generates a citation-traced .docx proposal — using Mistral OCR, an LLM, and hybrid pgvector + BM25 retrieval.

Bookkeeping teams re-key data from thousands of invoices in different layouts and languages, and label-based parsers fail when fields have no consistent labels.

Built a schema-driven multimodal extractor: each invoice page is rendered to an image and passed to Mistral Large with a Pydantic schema enforced via instructor, backed by a regex validation layer, then written straight to Excel.

State institutions require advanced AI capabilities but cannot risk exposing classified or citizen-sensitive data to external providers.

Engineered a fully sovereign AI stack deployed inside restricted government networks, enabling large-scale ML and LLM workloads without any external data dependency or leakage vectors.

High-stakes AI deployments fail when models produce confident but incorrect outputs under uncertainty.

Designed a multi-layer reasoning and verification architecture where model outputs are recursively challenged, cross-validated, and stress-tested before release.

Large-scale logistics environments suffer from human navigation inefficiencies and suboptimal routing.

Developed an adaptive ML-driven control system for autonomous agents that continuously learns spatial dynamics and optimizes retrieval paths.

Small teams were juggling their operations across spreadsheets and disconnected tools, with no single source of truth.

Built a full-stack multi-tenant SaaS platform with authentication, role-based access, billing, and real-time dashboards — everything a team needs in one place.

Large-scale search systems face latency and cost explosions as model complexity increases.

Redesigned the model architecture and inference pathways for a large-scale search platform to drastically reduce computational overhead while preserving ranking intelligence.

Manual interpretation of aerial and satellite imagery creates critical delays in conflict or disaster response.

Implemented a computer vision pipeline capable of large-scale structural anomaly detection and automated geospatial damage classification.

Critical intelligence signals are buried inside massive volumes of unstructured voice communications.

Built an ML-based speech and entity extraction system to automatically surface names, locations, and operationally relevant references.

Rapid infrastructure changes and unauthorized developments are difficult to monitor at scale.

Designed a vision-based temporal comparison system highlighting structural and environmental deviations across satellite datasets.

A service business was losing revenue to double-bookings and manual scheduling done over phone and paper.

Built a real-time booking platform with live availability, instant notifications, and an admin panel — bookings sync across every device the moment they happen.

Municipal departments are overwhelmed by unrouted citizen inquiries and manual document processing.

Automated triage system using custom BERT-based intent classifiers to route citizen requests and pre-extract entities from application forms.

Onboarding and policy retrieval across global retail units are fragmented across disconnected legacy document stores.

Developed a centralized RAG assistant indexing 10k+ internal documents with hybrid retrieval and semantic reranking.

Unplanned factory downtime due to critical equipment failure costs millions in annual lost production.

End-to-end unsupervised anomaly detection pipeline combining vibration analysis and thermal modeling to predict failure windows.

Manual property appraisal is subjective and fails to scale across high-volume portfolio acquisitions.

Automated valuation system using multi-agent scrapers and hedonic regression models to provide data-backed pricing at scale.

Legacy OCR systems fail on non-Latin scripts and non-standard document layouts like handwritten invoices.

Combined DocTR vision models with LLM reasoning to extract structured data from Arabic, Cyrillic, and Latin document sets.

Signature-based security fails to detect sophisticated lateral movement and novel zero-day attack vectors.

Trained Graph Neural Networks (GNNs) on network flow data to identify statistically significant deviation patterns in node communications.

Sharing sensitive financial or medical datasets for model development is blocked by stringent privacy regulations.

Developed a GAN-based generator that produces synthetic tabular data preserving statistical distributions while ensuring DP guarantees.

Local dispatch offices tracking multi-vehicle fleets via manual synchronization, leading to critical routing errors.

My initial production-grade fullstack deployment. Built a real-time tracking and manifest generation system for logistics operators.

Massive relational tables (100M+ rows) suffering from query timeouts during peak reporting hours.

Optimized complex join logic and implemented materialized view partitioning to reduce computational load on core databases.

Heavy reliance on bloated frameworks leading to poor performance on low-end hardware.

Engineered a custom micro-framework in Vanilla JS with a diff-based DOM update engine. Achieved near-instant TTI.

Competitive market data collection blocked by sophisticated anti-bot countermeasures and IP rate-limiting.

Built a distributed scraping cluster with automated proxy rotation and headless browser orchestration to harvest data at scale.

Most developer portfolios look the same. I wanted something that proves I can mix bold design with the complex engineering that runs behind the scenes.

Designed and built this site end to end: a retro pixel aesthetic, 3D scenes with React Three Fiber, GSAP and Framer Motion animations, and a custom WebGL/ASCII renderer — all on Next.js and deployed on Vercel.