Vasile Bogdan Godja

PrimărIA — multi-tenant tax SaaS for Romanian municipalities

Public GovTech portfolio implementation with tenant isolation and deterministic workflows.

Active-development tax administration platform with deterministic calculations for buildings, land, and vehicles; role-scoped portals; audit records; XML export tooling; and offline-oriented PWA support. It is portfolio and pilot-stage work, with no public deployment or production/compliance certification claimed.

Grant Scout — cited EU funding agent demo

A deliberately small agentic-RAG system whose evidence and control flow can be inspected end-to-end.

Searches a curated 30-record snapshot dated 29 August 2026, applies deterministic eligibility checks, and cites official sources. CI runs lint, typecheck, tests, and build. Dated model-backed evaluation reports separately record tool use, retrieval, citations and failures. Deployment is pending; this is not a live funding catalogue or funding advice.

EuFund — AI-assisted EU funding workflow

An active-development portfolio prototype for deterministic, auditable funding workflows.

Discovery → eligibility → proposal drafting → compliance review, with structured outputs, citation enforcement, and cost-aware multi-provider LLM routing. The source repository is private and there is no public or live deployment; the case study remains available as design and implementation context.

MitchAI — hospitality operations AI workbench

Active-development portfolio work; no live or production-service claim.

The primary implementation is a Node.js/TypeScript/Express backend with a Next.js frontend and PostgreSQL persistence. Local verification: 193 backend tests pass, with none skipped. Eight n8n workflows execute through 10 synthetic HTTP scenarios, including a simulated provider failure. Provider endpoints and the database sink are substituted; the combo route returns a plan. Public-history cleanup and current production verification remain pending.

AI Gateway — unified multi-provider LLM middleware

OpenAI-compatible router for OpenAI, Anthropic, Gemini, xAI, Groq, and Perplexity.

A multi-provider gateway consumed by PrimărIA’s fiscal assistant. Published source includes token authentication, rate limiting, spend controls, structured logging and response validation for selected JSON endpoints. Tests exercise routing and failure handling; mocked health checks do not verify real provider execution.

Offline portfolio project. Source and tests are public; a hosted demo is intentionally not offered. Salt & Standard is the live business site.

Local AI Gateway — private knowledge-base service

FastAPI, hybrid retrieval, document ingestion and an MCP bridge.

Inspectable local-model infrastructure with authenticated file access and destination checks on web redirects. 83 tests, including security regressions. No hosted service or new real-model quality benchmark is claimed.

Local LLM MCP — inference queue and deadline handling

A focused local-model integration with explicit failure behavior.

MCP tools over a local OpenAI-compatible model endpoint, with queueing, deadlines and telemetry. 31 tests use local mock servers; benchmark reports are historical measurements rather than a guarantee for another machine.

eufund-exhibit — phase-gated agent runtime for EU-funding workflows

Public exhibit extract from the EuFund prototype: write gates, cited retrieval, tiered routing.

A publicly inspectable slice of the private EuFund agent runtime — phase-gated execution, authorization checks for controlled write tools, citation-enforced retrieval, and tiered LLM routing by task cost. Runs with synthetic fixtures and injected providers; 57 tests and CI. A sanitised Cloud Build reference documents the private platform’s deployment design; the exhibit excludes its database and app.

Japan_Agent — human-approved research agent

Paper-first research infrastructure with proposal approval and an audit ledger.

Structured research is separated from order execution. Immutable proposal tickets, Telegram approval, fresh-data checks and a hash-chained ledger govern the execution path. Paper/research software; no live-trading permission or verified track record is claimed.

review-responder — review → draft → approve → publish pipeline

Human-in-the-loop response drafting for hospitality guest reviews.

Pulls guest reviews, drafts an on-brand reply, holds it for human approval, then publishes through an explicit approval workflow. The repository documents publication state, reconciliation and tested failure cases; no live Google Business Profile publication is claimed.

Veloce Transfers — deterministic fare engine

Rule-based pricing engine for a private transfer operator.

Deterministic fare calculation across zones, vehicle classes, and surcharges for a private transfer business — built to be auditable rather than a black-box quote. 86 tests.

Benjamins at the Landsby — 13 hospitality operations tools

Runnable operations demonstrations built around a restaurant-management application.

Thirteen focused tools for reconciliation, briefing, forecasting and reporting, with synthetic sample data and inspectable outputs. 173 tests. The repository demonstrates software behavior; restaurant adoption and business results are not claimed.