Applied RAG research
RAGnosis
Applied RAG research benchmarking how retrieval strategy, not model choice, drives question-answering accuracy over a synthetic clinical database.
A resource-constrained study run on local hardware that pits a baseline vector-search pipeline against an advanced one built from query rewriting with dual retrieval, small-to-big parent expansion, precomputed aggregate rollups, and cross-encoder plus listwise reranking. Across eight tracked runs it isolates what actually moves accuracy by question type, factual, numerical, and holistic, and ships an interactive benchmark report with a leaderboard, a question bank, and reproducible run configs. The headline result: changing what the system retrieves helped more than changing the model.
Built withPythonOllamaChromaHugging FaceJina Reranker v3BGE RerankerOpenAI EmbeddingsRAG Evaluation
FocusRetrieval strategyQuery rewritingRerankingSmall-to-big expansionRAG benchmarking
Synthetic data and local LLM tooling
data-alchemy
Local-LLM synthetic data generator that turns a dataset brief into related table schemas and downloadable CSV, JSON, and Markdown exports.
Built as a Python and Gradio workflow around Ollama's OpenAI-compatible API, Faker, and notebook-friendly execution. It shows applied work on prompt design, schema planning, generation constraints, and reusable data creation for RAG demos, evaluation sets, and product prototypes.
Built withPythonGradioOllamaOpenAI-compatible APIsFakerJupyter NotebookSynthetic DataCSV/JSON Export
FocusDataset generationSchema designLocal model orchestrationRAG evaluation dataPrompt constraints
Multimodal local AI assistant
zyra
Plant-care assistant that combines image understanding, speech input and output, local GGUF inference, and botanical tool lookup.
Built as a Gradio app with llama.cpp, quantized Gemma vision-language models, Whisper, Kokoro, and optional Trefle API lookup. It demonstrates the parts of an AI assistant that matter in production: multimodal input, local inference tradeoffs, tool-backed answers, and a user-facing interaction loop.
Built withPythonGradiollama.cppGemma GGUFWhisperKokoro TTSTrefle APITool CallingMultimodal AI
FocusLocal vision-language inferenceSpeech-to-textText-to-speechExternal API toolsAssistant UX
Full-stack AI chat product
pretendo
Production-minded AI chat app with a Next.js client, FastAPI service, OpenRouter model routing, browser SQLite, and Vercel deployment.
Built as a multi-service monorepo with clear API boundaries, model validation, backend-owned prompts, request limits, provider timeouts, local chat persistence, analytics, and focused tests. It presents the engineering shape of a small shipped product rather than a throwaway demo.
Built withNext.js 16React 19TypeScriptTailwind CSS v4FastAPIPydanticOpenRoutersql.jsVercel ServicesAPI Security
FocusModel routingAPI boundary designLocal chat persistenceRequest validationProduction deployment
