Postponed experiment in content-based PII classification using a small multilingual encoder per entity type. The architecture and evaluation plan exist; training, validation, packaging, and publication are paused.
~270Mplanned model size per entity
90%+ targetper-cohort evaluation gate
CPU-firstintended deployment constraint
Architecture notes
- Content-aware hypothesis: classify actual cell values rather than infer PII from column names.
- Specialized models: test whether smaller per-entity encoders outperform a general 7B LLM on structured classification.
- Honest evaluation: report precision, recall, and F1 per cohort, with an explicit ambiguous band and no aggregate-only success claim.
mmBERTONNX int8Multilingual NLPData Privacy
Self-hosted bilingual household system for shopping, pantry, recipes, receipts, and spending. Its unusual part is a fully local OCR + VLM + visual-RAG pipeline tuned for Georgian receipts on one 16 GB consumer GPU.
0 cloudlocal inference pipeline
6 models / 1 GPUexplicit swap windows
~3-7 minreceipt processing
Engineering notes
- Multi-turn re-read: the VLM revisits partial receipt regions instead of forcing a single-shot answer.
- Positive and negative visual memory: pgvector-backed embeddings learn both canonical products and crops that should always be ignored.
- Hardware-aware orchestration: explicit model windows keep interactive paths resident while receipt jobs pay the swap cost.
FastAPIPostgreSQLpgvectorOllamaVisual RAGWhisper
Research into whether quantized 8B models can sustain model-driven behavior without an agent framework supplying the loop. Rev 1 produced an apparent signal; manual review showed mostly repeated game-shaped filler, so every rev 1 conclusion was retracted.
2,680 runsrev 1 corpus, findings retracted
0 LLM gradersmechanical outcomes only
300 plannedrev 2 test matrix
Method notes
- Retractions on contact: the strongest-looking trace failed a close read, so the finding was removed rather than rationalized.
- Pre-committed metrics: topic distance, novelty against prior turns, and frame-break detection are defined before new runs begin.
- Rev 2: a silence-as-continuation protocol removes the repeated "what next?" prompt that makes the framework itself the source of agency.
Llama 3.18B QuantizedRAGEmbeddingsReproducibility
Parked design for an audit layer between AI agents and warehouses. The concept records agent identity, SQL operations, touched tables and columns, and policy decisions without pretending the unbuilt system already exists.
49 stepsversioned implementation plan
~300 LOCfirst milestone target
Parkeddesign retained, build paused
Design notes
- Audit boundary: intercept data operations close to the database rather than infer them later from LLM traces.
- Transparent integration target: SQLAlchemy, direct clients, and MCP middleware feed one structured event model.
- Policy path: the same interception point could later block disallowed tables or operations before execution.
PythonsqlglotMCPData Governance