ProjectsLLM Fine-Tuning & Local InferenceQwen2.5-Coder 0.5B logo model with LoRA experiments, a custom PyTorch training loop, GGUF quantization and CPU inference. Recorded CPU benchmark: 24/24 valid outputs, 21/23 brand names preserved, 26.2 s mean generation time. Small synthetic-data experiment. Tools: Qwen2.5-Coder · PyTorch · LoRA · GGUF · CPU inference Multilingual Career ModelLocal model trained on CV and career Q&A, with translation before and after inference, output verification, held-out evaluation and hallucination checks. Recorded multilingual regression: 42/54 (77.8%) against a 90% gate. Not promoted; fine-tuning experiment without RAG. Tools: Local LLM · Fine-tuning · Translation · Evaluation AI-Assisted DevelopmentClaude Code workflow with Jira tasks, isolated Git worktrees, per-task environments, browser verification, a UAT gate, layered instructions and code-graph support. Public guide and templates for verification gates and seven-language AI-text quality checks; environment-specific ws CLI is not included. Tools: Claude Code · Jira · Git worktrees · Browser verification · UAT AI Benchmark & Market IntelligenceCollection and normalization pipelines turn model accuracy, model pricing and cloud GPU pricing data into comparison views. 15 benchmark sources · 12 GPU pricing sources · 3 model-pricing sources. Tools: Data pipelines · AI benchmarks · GPU pricing Resilient Web Data PipelinesBright Data and direct HTTP/cURL scraping workflows with retries, honeypot detection and faulty-data checks before downstream use. 500+ URLs processed per month. Tools: Bright Data · HTTP · cURL · Data validation Multilingual Content & ReviewsTranslation infrastructure for articles and static content, multilingual AI-text quality checks, and review analysis for spam and human-review prioritization. 500+ articles and static content across seven languages. Tools: Translation · Multilingual pipelines · Review analysis Private Codebase RAGFully local Ollama, DeepSeek-R1 and ChromaDB setup for querying PHP code and database schemas. 50,000+ lines of PHP · about 40% less manual inspection time. Tools: Ollama · DeepSeek-R1 · ChromaDB · PHP · RAG SüperLig Data ExplorerLive Next.js application with source-linked match records, season and team filters, and an explicit decision-impact scoring model. Three seasons · 1,028 match records · 4,754 card events. Suspension and VAR research remains incomplete; scores are not evidence of referee bias. Tools: Next.js · TypeScript · Data validation · Source provenance PHP Coding-Agent BenchmarkTask-based evaluator for PHP coding agents covering snapshot isolation, idempotency, persistence, concurrency and crash recovery, with static analysis and mutation scoring. Five task variants with recorded runs and calibration controls. Small run counts limit comparisons between closely scoring models. Tools: PHP · Python · PHPStan · Mutation testing · Agent evaluation Rock Paper Scissors with Workers AIBrowser-based game that sends webcam frames to Cloudflare Clef for bounded rock, paper and scissors classification. Client-side motion gating reduces unnecessary requests; a Worker combines Durable Object daily budgets, per-IP quotas and burst rate limiting. Public playable demo with server-side AI inference and usage controls. Camera access requires HTTPS or localhost; classification depends on the model and camera conditions. Tools: Cloudflare Workers AI · Clef · Durable Objects · JavaScript · Camera API Hook Bench — AI Design & FEA BenchmarkExperimental benchmark for AI-designed J-hooks created through Blender MCP. Python validates STL geometry against shared constraints and runs 2D finite-element analysis, with reference hooks, print profiles and a documented physical-test protocol. Repository includes two AI design runs, calibration geometry and FEA predictions. Physical breaking-load measurements are not yet recorded; predictions are not measured strength, and 2D analysis omits layer adhesion and large deformation. Tools: Python · Blender MCP · STL validation · scikit-fem · 3D printing |