§ projects
Open-source systems & research artifacts.
A selection of open-source systems and research artifacts by Ademar Tutor: GLAMLI (Master's dissertation), LexiTau and LexExtract (agentic Text-to-SQL over financial documents), an LLM-as-a-Judge evaluator, and SafeBeautyLedger. Most started as research questions and ended up as code — from papers, from production problems, or from a weekend's curiosity. Each links to a write-up, the repo, or both.
- GLAMLI (Dissertation) · in progress
- LexiTau · agentic Text-to-SQL
- LLM-as-a-Judge Evaluator · evaluation methodology
- LexExtract · OCR → Text-to-SQL
- SafeBeautyLedger · blockchain traceability
-
GLAMLI (Dissertation)
Guided LLM-Augmented Machine Learning Interface — an agent that walks non-experts through a real ML workflow (load → profile → train → evaluate → deploy).
- Planner-with-revision orchestration topology
- Typed tool registry (data primitives) the agents call into
- Sandboxed code execution for generated Python
- Structured observability traces for trust calibration and failure analysis
repo link when public -
LexiTau
Agentic Text-to-SQL pipeline for financial document querying. End-to-end: profiling → metadata → schema linking → SQL draft → safety checks → execution → narrative explanation.
- RAG-based schema context retrieval (Jaccard over metadata, cosine on queries)
- Multi-tenant isolation: per-tenant embeddings, scoped auth, row-level policies
- OCR ingestion (Azure Form Recognizer) normalised into tenant-scoped tables
- Stack: FastAPI · PostgreSQL+pgvector · SQLAlchemy · Celery/Redis · Next.js · Docker Compose
Multi-tenant RAG with real isolation is rare in public artifacts.
-
LLM-as-a-Judge Evaluator
Statistical-rigor evaluation framework for LLM-judged outputs. Seven SPIN-based scoring dimensions, evaluated with QWK, Pearson r, and ±1 accuracy — three metrics that disagree about which judge is best, which is the point.
- Pydantic schema enforcement; prompt-level contracts for deterministic, auditable outputs
- Tenant-scoped prompt versioning with audit trails (Postgres + SQLAlchemy)
- Score-once / evaluate-many workflow that cuts LLM cost without losing fidelity
- Multi-LLM provider integration (OpenAI · Anthropic · Google)
- Stack: FastAPI · PostgreSQL · SQLAlchemy · LangChain · Docker · Azure Container Apps
-
LexExtract
OCR-to-Text-to-SQL platform for financial document querying. Natural-language queries over PDF-extracted financial data.
- Full pipeline: PaddleOCR → LLM-based Text-to-SQL with memory-efficient processing and progressive fallback
- Explainable query outputs through structured SQL generation and validation layers
- Stack: FastAPI · PostgreSQL · PaddleOCR · Mistral 7B · Docker
-
SafeBeautyLedger
Blockchain traceability for beauty-industry product provenance. Immutable history, public verification, on-chain product registry.
- Solidity smart contracts (Hardhat) with versioned product records, event-driven history retrieval
- On-chain / off-chain data flow integration
- Stack: Solidity (Hardhat) · Node.js + TypeScript · PostgreSQL · Ethers.js · Next.js
Outside the multi-agent work, but a useful counterpoint — different problem class, same engineering discipline.