Document ingestion
RAG pipeline ready
$ ingest workspace_docs
→ chunk · embed · index · verify sources
Alexandru-Stefan Cervinschi · AI-native product builder
Senior Full-Stack Developer & AI Engineer
5+ years building production web and mobile systems. I build software and AI-powered products end to end — system architecture, APIs and user interfaces, LLM workflows, infrastructure, deployment and observability.
Open to remote opportunities across Europe · B2B available through Romanian SRL
Based in Romania · Europe timezone
contact@hyperdevelopment.ro · English · Romanian
Agent workflow
running
tool retrieve_context(question)
tool draft_answer(schema)
check source coverage + tone
Validated output
sources verified
answer grounded_response.v2
sources policy.pdf · faq.md · admin note
TypeScriptNestJSNext.jsReact NativeFastAPILangChainLangGraphRAGpgvectorPostgreSQLDockerKubernetesAWS
01 / Selected Work
Products and tools I've designed, built and shipped.
Document-grounded customer support, repository-aware code review and exam preparation — each one built from the database schema up.
Nexivo
Multi-tenant AI support, grounded in each customer's own documents.
A SaaS platform where every workspace ingests its own documents into a pgvector store and answers customer questions through an embeddable widget. LangChain and LangGraph agents handle retrieval and answer drafting, streamed token by token, with source attribution on every response. The platform side covers tenant isolation, role-based access, Stripe subscriptions with usage quotas, and tracing over each request. Live, with paying customers.
- Multi-tenant architecture with RBAC
- RAG over customer documents (pgvector)
- LangGraph agents, streamed responses
- Stripe subscriptions and usage quotas
- Request tracing and answer feedback loop
- Embeddable support widget
Next.js · NestJS · TypeScript · PostgreSQL + pgvector · Prisma · LangChain · LangGraph · OpenAI API · LangSmith · Stripe · Docker
Interface preview
Can I return a product after the standard window?
Based on the uploaded policy, exceptions can be reviewed when the order has documented delivery delays or a warranty-related issue. The support team should request the order ID before confirming next steps.
source found · returns-policy.pdfMedizi
A daily practice system for the Romanian medical school entrance exam.
A study app for students preparing the UMF Iași admission exam, built around a daily plan rather than a question bank — mastery tracking, diagnostics and spaced review decide what comes next. Question content is generated through a provider-routed AI layer with caching and per-request cost limits, then has to clear an editorial review gate before any student sees it. React Native app on a NestJS API, with a Next.js admin panel for the editorial side.
- Daily plan and mastery tracking
- Diagnostics and spaced review
- Provider-routed AI with cost guardrails
- Editorial review gate before publish
React Native (Expo) · NestJS · TypeScript · PostgreSQL · Prisma · Next.js · Anthropic + OpenAI · AWS S3 · Sentry
Product in development
Daily learning planEditorial review
Cell biology foundations
Chemistry problem set
Concept explanation draft
Adaptive practice queue
AI PR Reviewer
Repository-aware review, beyond the diff.
An AI-assisted GitHub PR reviewer that retrieves relevant repository context before analyzing code and generating review feedback. Repository files are embedded and searched semantically, the model returns structured findings with severity and rationale, and secrets are redacted before anything leaves the repo. OAuth2 for GitHub access, per-review token tracking, deployed with Docker.
Stack details configurable after repository URL is set.
backend/app/retriever.py
backend/app/llm_review.py
README.md
frontend/src/app
@@ repository context @@
+ retrieve related files before scoring
! validate structured output schema
+ attach severity and rationale
02 / Capabilities
What I own when I join a product.
Full-Stack Product Delivery
Frontend, APIs, database design, authentication, billing, CI/CD and release — I take a feature from schema to deployed, and stay responsible for it afterwards.
RAG & Retrieval Systems
Ingestion pipelines, chunking and embeddings, pgvector search, and answers grounded in sources the user can check.
LLM Agents & Workflows
LangGraph and ReAct agents, tool calling, structured outputs, and tracing to see what the model actually did.
Backend & Infrastructure
NestJS and FastAPI services, REST API design, PostgreSQL schemas, multi-tenancy, OAuth2/JWT and RBAC — running on Docker and Kubernetes with CI pipelines and observability in place.
AI-Assisted Engineering
Claude Code and Codex are part of my daily workflow, with review and test coverage around whatever they produce.
03 / Stack
What I work with.
Technologies I've used in real projects — not a list of things I've only read about.
AI / LLM
Backend
Frontend
Infrastructure
Data
Engineering
AI-assisted development
04 / Approach
Build. Validate. Improve. Scale only what works.
01
Understand the workflow
I start from the real user problem and the constraints around it, not the AI feature.
02
Build the smallest useful version
A working slice in front of users early, so feedback comes from usage instead of opinions.
03
Learn from what breaks
I read the failures, the traces and the support questions before adding more surface area.
04
Harden what proves valuable
Security, reliability, observability and performance work goes where the usage actually is.
05 / About
Alexandru-Stefan Cervinschi.
Founder of Hyper Development
Senior Full-Stack Developer & AI Engineer
I've spent five years as a software engineer on production web and mobile products — owning features from database schema through APIs to the deployed frontend, across NestJS and Rails services, React and React Native clients, PostgreSQL, and the CI/CD around them.
More recently I moved that same ownership into AI products: retrieval pipelines over customer documents, LangGraph agents with structured outputs, and the tracing and evaluation needed to keep them honest once real users depend on them. Nexivo is where most of that work lives.
I work through Hyper Development ASC SRL, my Romanian company, when a B2B contract is the simpler setup for both sides.
Based in RomaniaEurope timezone
Remote across EuropeFull-time or contract
Experience5+ years in production teams
LanguagesEnglish · Romanian
06 / Contact
Let's talk.
I'm open to Senior Full-Stack, AI Engineer and GenAI Engineer roles, remote across Europe. Happy to walk through the architecture of anything above in detail.
This check depends on repository context. Add a fallback path when retrieval returns no matching files.