Projects

AI engineering, AI architecture, and integration with strong engineering discipline

Problem

A portfolio backend with AI functionality should remain available even when a provider hits quota limits or is temporarily unavailable.

Solution

FastAPI with LangChain multi-fallback: Google Gemini, OpenAI, and Groq are used depending on feature and availability. The goal is reliable downstream processing of LLM results via validated, typed data structures (Pydantic structured output) for the demo endpoints.

Result

CV assistant, vision analysis, sentiment analysis, and agent chat live at sinanucar.com — with resilient operation even during provider outages.

Technical details and decisions

Technical Architecture

Astro frontend with FastAPI backend, four demo endpoints, separate fallback chains for text and vision, and Pydantic models for structured responses.

Engineering Decisions

Model text and vision fallbacks separately
Use Pydantic structured output instead of manual JSON parsing
Portfolio demo with constrained inputs and no permanent chat history storage

Demonstrates

Controlled assistant interaction, structured text classification, multimodal analysis, and resilient provider fallback.

Status

Live proof of controlled AI integration on sinanucar.com.

FastAPILangChainPydanticAstroTailwind CSSRender

View AI integration →

Problem

Self-managed homeowner associations with 3–20 units organize themselves via email threads, spreadsheets, and paper minutes. At the same time, German WEG law (§24, §25, §28) demands correct, permanently traceable resolutions, votes, and claims.

Solution

A multi-tenant platform built as a modular monolith: the Next.js 16 web app and the FastAPI agent service both talk to Supabase Postgres with the user JWT. Security and legal rules live as database contracts: RLS tenant isolation, an append-only resolution register, an HMAC-chained audit log, and trigger guardrails that block direct agent writes to legally binding data.

Result

A verified working state with web app, agent service, and 59 documented migrations; core rules are secured as pgTAP-tested database contracts, and the project state is machine-checked for freshness.

Technical details and decisions

Technical Architecture

Modular monolith with two deployments: Next.js 16 (server components, server actions) and FastAPI with LangGraph state machines; Supabase Postgres with RLS, auth, storage, and vault in Frankfurt; type safety across service boundaries via OpenAPI codegen.

Engineering Decisions

Security as a database contract: RLS, trigger guards, and least privilege instead of app-level promises
AI as suggestion only: database triggers block agent writes to votes, resolutions, and minutes
Immutability over recalculation: append-only audit chain and versioned, materialized dues records

Demonstrates

Multi-tenant architecture, database security (RLS, pgTAP), AI guardrails with human-in-the-loop, and domain-driven schema decisions along German WEG law.

Status

Local development project with a Supabase cloud backend; no hosting, no billing, no customer operation.

Next.js 16FastAPILangGraphSupabasePostgres RLSpgTAPTypeScriptPython

View the tech deep-dive (German) →

Problem

Company knowledge often sits across scattered local documents, status data, and decisions. AI answers are only useful when search scope, sources, permissions, and evidence status are correct.

Solution

BetriebsGehirn starts with the knowledge infrastructure: local source management, parsers, chunking, embeddings, vector index, source status, document store, and separate setup/health states.

Result

A RAG-oriented working state that treats answer quality, source confidence, and honest non-answers as core product concerns.

Technical details and decisions

Technical Architecture

Next.js app with modular core/module/adapter layers, local JSON stores, persistent vector index, optional LLM provider, and API endpoints for knowledge, sources, documents, setup, and health.

Engineering Decisions

Knowledge infrastructure before chat UI
Source status, source scope, and evidence assessment as architecture concerns
Adapters instead of hard-wired providers, databases, or parsers

Demonstrates

RAG architecture, local knowledge processing, retrieval quality, source grounding, and realistic AI boundaries.

Status

Local development project; not a finished SME product and not running customer data.

Next.js 16TypeScriptRAGEmbeddingsVector IndexLocal LLMAdapters

Read the case note →

Problem

Speech therapy documentation is professionally demanding and time-consuming. Structured reports are mandatory, but writing them manually consumes time that is missing from the actual therapy.

Solution

The prototype combines transcription, structured extraction, and validated report generation. Focus: traceable draft reports, clear user guidance, robust data structuring, and professionally useful outputs.

Result

A traceable AI workflow that converts unstructured speech input into professionally structured, validatable draft reports.

Technical details and decisions

Technical Architecture

Frontend audio upload, FastAPI pipeline for transcription and report generation, and Pydantic models for validated structured outputs.

Engineering Decisions

Structured output before free text so domain users can review and edit draft reports
Clear separation between AI suggestion and professional control
Compact prototype instead of oversized platform logic

Demonstrates

Transcription, structured extraction, validated report generation, and domain-aware user guidance.

Status

Prototype / portfolio project; no production processing of health data.

WhisperFastAPIPythonLangChain/LangGraphPydanticStructured OutputsRAG

View demo →

Problem

Starting from a collaborative platform concept for creative and film projects: how can authentication, data modeling, collaborative product logic, and AI functionality be combined so that LLM outputs flow into real application processes in a traceable way?

Solution

RealizeTogether serves as a testbed for AI integration: authentication, data model, FastAPI backend, Supabase, LangChain, and Pydantic structured outputs. The project demonstrates validated AI outputs, workflow integration, role/data logic, and safe extensibility.

Result

Ongoing R&D/MVP platform with authentication, data layer, and AI pipeline — a basis for evaluating AI integration in product workflows.

Technical details and decisions

Technical Architecture

Next.js frontend, Supabase for authentication and data storage, FastAPI as AI backend, LangChain and Pydantic for validated LLM outputs.

Engineering Decisions

Keep AI complexity isolated in the backend
Do not delegate role and data logic to the LLM
Use validated outputs as the interface between LLM and product workflow

Demonstrates

Workflow integration, auth/data model, safe extensibility, and structured AI outputs.

Status

In development; R&D/MVP, not a finished enterprise product.

Next.js 16SupabaseFastAPILangChainTypeScriptPython

View demo →

Problem

AI coding tools can produce output quickly, but without shared project truth, handoffs, and review gates they lose continuity.

Solution

CoWork-OS organizes agentic software development as a local Markdown control plane with roles, workflows, rules, adapter thinking, memory, project files, and review stops.

Result

A structured working mode that treats AI tools not as magic, but as part of traceable software development.

Technical details and decisions

Technical Architecture

Git-versioned Markdown structure with AGENTS rules, workflows, memory, reviews, project files, public/private boundary, and tool-agnostic adapters.

Engineering Decisions

Repository files instead of fleeting chat history as project truth
Controlled autonomy with clear approval and review boundaries
Tool- and model-agnostic structure instead of tying the workflow to one AI tool

Demonstrates

Agentic development, multi-agent handoffs, review gates, project governance, and AI-assisted work organization.

Status

Local control-plane/workflow system; not a fully automated enterprise platform.

MarkdownGitCodexClaude CodeGeminiAgentic Workflows

Projects

This Portfolio

WEG-Verwaltung

BetriebsGehirn

Speech Therapy Report Agent

RealizeTogether

CoWork-OS