11/11/2026 updated
40 % partially available
AI Architect · Co-founder and CTO at Inflexa · Agent runtimes, LLM cost, and security for AI agents
Cluj-Napoca, Romania
Only remote
MSc Information and Computing System Security (TU Cluj-Napoca), BSc Computer Science (UBB Cluj)About me
I build AI agents for the day after the demo: a run that crashes at hour three, an LLM bill that doubles, and an agent that acts for a user. Two-time co-founder and CTO. I halved our production LLM spend, and our agent runtime is open source. Part time, 8–20 hours a week.
AI AgentsArgo CDGitOpsGitHub ActionsOpenTelemetryHelmArtificial IntelligenceAmazon Web ServicesGoogle BigQueryElasticsearchPython (Programming Language)PostgreSQLNode.jsSoftware ArchitectureTypeScript
AI agents and agentic AI systems
- Architecture of multi-hour AI agents: durable execution (DBOS workflows on PostgreSQL), crash recovery, retries that never run a command twice, budgets on every model call.
- Sandboxes for AI-written code (gVisor, no network egress).
- LLM cost control: prompt caching, append-only agent conversations, LLM gateways (routing, fallback, cost attribution). Halved a production LLM spend at nearly the same token volume.
- Programmatic tool calling, tool design, evals with deterministic graders, prompt hill-climbing without regressions.
- Delegated authorization for agents that act for users: signed passports, revocable scoped mandates, short-lived service tokens.
- Models: Anthropic Claude, OpenAI-compatible APIs, self-hosted models (Qwen).
Cloud platform and security
- AWS: Organizations with separate accounts, EKS with Karpenter, RDS PostgreSQL, SSO, no static access keys, AWS Well-Architected review (all high-risk findings removed).
- Kubernetes, Helm, ArgoCD (GitOps), OpenTofu/Terraform, GitHub Actions with OIDC.
- Policy as code in CI: checkov, Trivy, KubeLinter, Kubescape. Signed releases (cosign, SLSA provenance, SBOM).
- Observability: OpenTelemetry, Grafana, alerts and dashboards as code.
- Mutual TLS, access-control models, audit logs. MSc in Information and Computing System Security.
Languages and backend
- TypeScript, Node.js, Go, Python, PostgreSQL, Elasticsearch, BigQuery, Google Cloud, Docker, microservices, API design.
- Spec-driven development with coding agents, AI review in CI.
Industries
- AI/ML, fintech and payments (fraud detection, 3-D Secure), B2B SaaS, cloud services.
Experience
Professional experience5 - 10 years
Freelancer since2022
Languages
EnglishFluentRomanianNative speaker
Project history
Inflexa runs AI agents for computational biology analysis. I own the architecture, the agent runtime, the cloud platform, and security, and I write much of the code.
- Took Inflexa's AI agent platform from the first commit to the first customer pilots in 10 weeks.
- Designed and built the agent harness behind it. Compute-heavy bioinformatics analyses run for hours and keep going through crashes and restarts without redoing work. AI-written code runs in isolated sandboxes. The harness is open source.
- Halved our production LLM spend at nearly the same token volume by redesigning how agents use the prompt cache. Programmatic tool calling then cut input tokens by about half in a measured research task.
- Put every agent under evals. A prompt change ships only when the evals pass with no regressions.
- Designed how agents act for users: revocable, scoped delegation with short-lived service tokens and no shared secrets.
- Wrote nearly all of the AWS platform: EKS, OpenTofu, and GitOps with ArgoCD. No static access keys: people use SSO, CI uses OIDC.
- Cleared every high-risk Well-Architected finding in three weeks, in about 70 hardening PRs.
Panomics is a multi-omics analytics platform that pharma companies license. I built it from zero as co-founder and CTO: Go and React, containerized analysis jobs on AWS Batch and in a partner cloud, a hierarchical access-control model, licensing over mutual TLS, and Stripe billing. A top-10 pharma customer renewed twice and doubled its contract. Early AI features: natural-language query building, metadata extraction, and an AI copilot. For a cloud platform partner, I also built a data-access tool for Eli Lilly research teams and extended FDA substance-registration software (GSRS).
Machine-learning fraud detection for online merchants. On the payments team, I built the query engine behind the merchant analytics dashboards, with nested aggregations over Elasticsearch and BigQuery (Go, GCP, Kubernetes). On the 3-D Secure team, I worked on card-authentication systems that process card data, and I took part in the on-call rotation.