Multimodal Agentic Cooking AssistantCook Mantra
An application that turns an ingredient photo into complete recipes through multimodal and specialized LangGraph agents for extraction, dish suggestions, nutrition estimates, preview images, and recipe generation.
The local-first workflow requires users to confirm detected ingredients, suggested pantry staples, and manually added items before generation. Only the confirmed list feeds recipe creation, while missing, optional, and replaceable ingredients remain explicit.
Next.jsReactTypeScriptPythonFastAPILangGraphOllama
Applied RAG EvaluationRAGnosis
An eight-run benchmark comparing retrieval and reranking strategies over a synthetic clinical database. The project evaluates how query rewriting, small-to-big retrieval, aggregate rollup documents, and rerankers affect retrieval and answer quality.
The documented study uses 1,200 synthetic patients, 2,200 appointments, and 30 evaluation questions. It archives configurations and results for all eight runs, while clearly limiting the findings to this synthetic dataset and evaluation design.
0.406 → 0.792 retrieval MRRBest documented answer-quality score: 4.622/5
PythonChromaOllamaOpenAI-compatible embeddingsHugging FaceBGE RerankerJina RerankerEvaluation pipelines
Full-Stack LLM Routing ApplicationPretendo
A deployed Next.js and FastAPI LLM application with model routing, validated API boundaries, backend-owned prompts, provider timeouts, browser-side persistence, analytics, automated checks, and Vercel deployment.
The public monorepo separates the user interface from the model-provider boundary. The FastAPI service owns prompt construction, validation, request limits, and timeout behavior; the Next.js client owns interaction design and local-first chat history.
Next.jsTypeScriptFastAPIPydanticOpenRouterSQLiteVercel
Local Multimodal AssistantZyra
A multimodal plant-care assistant combining image understanding, speech input, speech output, local GGUF inference, and optional botanical-data lookup.
An additional experimental build exploring local inference trade-offs and a multimodal interaction loop rather than a deployed customer system.
PythonGradiollama.cppGemma GGUFWhisperKokoro TTSBotanical API
Synthetic Data GeneratorDataAlchemy
A local-LLM workflow that converts a dataset brief into related schemas and downloadable CSV, JSON, and Markdown data for prototypes, RAG demonstrations, and evaluation workflows.
An additional experimental build focused on schema planning, generation constraints, and reusable synthetic data-not meaningful public adoption.
PythonGradioOllamaFakerOpenAI-compatible APIs
