Role and current status
Solo developer. The assistant is used in working finance and accounting workflows. Client details and usage volumes remain confidential.
Selected work01 02 03 04 05 06 Get in touch
In use · Confidential project · 2026
Privacy-First AI Assistant
As the solo developer, I built retrieval, privacy controls, model routing and a Flutter client for a confidential finance and accounting assistant. This case study describes the engineering scope at a high level. Client identities, data and source code are confidential.
Privacy and model routing
RAG / CAG / KAG retrieval · Entity filtering · pgvector or sqlite-vec
Privacy layer
PII detection · Reversible pseudonymization · Encrypted context storage
Confidential tasks
Local models through Ollama
Policy-approved tasks
Cloud models
What it does
Problem and constraints
Support finance and accounting workflows while controlling how confidential information reaches language models. The system needed local execution for confidential tasks and cloud access only for policy-approved tasks.
Retrieval architecture
Built RAG, CAG and knowledge-graph (KAG) retrieval across 10 FastAPI services. A hybrid retrieval cascade and entity filtering select context, with pluggable pgvector or sqlite-vec storage.
Privacy controls and model routing
PII detection, reversible pseudonymization and encrypted context storage form the privacy layer. Confidential tasks run through local Ollama models. Cloud models are used only for tasks permitted by policy.
Client and file operations
Built the Flutter/Dart client and a file assistant with confirmation gates and reversible moves, giving users explicit control over file changes.
Verification
Enforced a 95% CI coverage threshold for the privacy layer. This is a test-coverage requirement, not a claim of 95% answer accuracy. Retrieval benchmark results and production adoption figures are not published.
Built with
PythonFastAPIFlutterDartpgvectorsqlite-vecOllama
Want to talk through the thinking behind it?
I'm always interested in useful product and platform conversations.