Python backend · GenAI · reliable systems

Alexandru
Cojocari

Python backend engineer focused on practical GenAI. Around 10 years of professional Python backend experience, commercial Java experience, and a portfolio of hands-on projects across AI integration, evaluation, data systems, and test automation.

I build focused demos that make engineering choices visible: safe interfaces, measurable behavior, reproducible data, and tests that explain what a system does—and where its limits are.

PythonFastAPIJavaGenAI integrationRAG & retrievalEvaluationData integrityPytest & Playwright

Selected work

Runnable portfolio demos exploring backend reliability, applied AI, data safety, and quality engineering.

ML evaluation · release safety

Compares champion and challenger model scores on the same labeled examples. Paired bootstrap intervals, required-slice gates, an empirical p95 latency check, and a replayable JSON decision make the release criteria inspectable.

Synthetic demo data. A CI decision aid for human review—not an automatic deployment or proof of model equivalence.
Python backend · GenAI integration

FastAPI demo covering request idempotency, persistent per-tenant quotas, provider adapters, bounded retries and fallback, plus a minimal request audit.

The included provider is deterministic and offline. Local portfolio demo, not a production gateway or billing system.
Document AI · validation

Extracts invoice and receipt fields from PDFs and images with pdfplumber, Tesseract OCR, rules, optional provider fallback, and Pydantic financial consistency checks.

Scoring uses a seeded synthetic corpus; the optional live API is not exercised in CI, so the accuracy table is not a production claim.
Data engineering · reproducibility

FastAPI service for schema-validated immutable snapshots, canonical SHA-256 fingerprints, manifests, and parent-linked dataset versions.

SQLite portfolio implementation with synthetic data; no authentication or tenant isolation. Not for confidential data.
AI security · policy engine

Rule-based HTTP firewall for text before or after an LLM call. Policies can allow, deny, or rewrite with a risk score and the rule that fired.

Offline pattern matcher; no model calls. Partial OWASP coverage, English phrase rules, and not a production WAF.
AI tools · data safeguards

Natural-language analytics over a small SQLite warehouse, with SQLGlot parsing, read-only checks, row caps, and query timeouts.

Default behavior is a fixed offline query catalog; the model endpoint is optional. Demo database, no auth or tenant isolation; not a production text-to-SQL system.
Retrieval · offline evaluation

Compares BM25, FAISS dense retrieval, and reciprocal-rank fusion on a project-authored FAQ corpus, with retrieval metrics.

Checked-in evaluation uses a deterministic hashing embedder; CI does not run the optional neural model. The corpus is not production query data.
QA automation · Python

Self-contained demo shop tested with httpx API checks and Playwright UI flows using pytest and Page Objects; CI captures failure artifacts.

The shop is part of the repository, not an external store. No payment is made; CI runs Chromium.
How I approach engineering

Practical systems, honest evaluation.

My main direction is Python backend engineering with GenAI and applied AI. I also bring commercial Java experience and a QA automation background, which shapes how I test boundaries and communicate system limits.

I use AI coding tools as part of development and carefully review and verify their output before treating work as complete.

These projects are independent portfolio demos. Synthetic or local data, mocked providers, and simulated behavior are called out on the relevant project cards and in each repository README.