Yegor Gaidar
Available · Q4 2026
Software engineer · Paris / remote, EU Open · Q4 2026

I build the system, then I keep it running.

Web applications, real-time interfaces, data pipelines, self-hosted models — and the infrastructure underneath all of it. I stay for the unglamorous half: alerts, backups, incidents, the page at 3 a.m. That is where a system actually lives.

Read the writing What I work on
Base
Paris, FR
on-site or remote, EU
Daily stack
TypeScript · Python · Go
Rust and Solidity where they earn it
Running
Proxmox · Docker · Cloudflare
self-hosted, no rented brain
Writing
22 engineering notes
Blog →
01 — What I work on

Four directions,
one habit: ship it, then operate it.

Most people hand a project over and leave. I stay for the part where it meets real traffic, real data and a real Monday morning — and that changes how the first half gets built.

01 Web applications & real timeBusiness apps people use all day Booking flows, back-office tools, dashboards, live interfaces. From schema to deploy: API, state, permissions, tests, the works.
  • Next.js · React · FastAPI · PostgreSQL
  • WebSockets, optimistic UI, offline tolerance
  • Delivered with tests and a runbook
 Next.js · FastAPI 02 Self-hosted AIModels on your hardware, numbers to prove it Small models that do one job well, retrieval over your own documents, distillation of a big model into a small one — measured on the hardware you already own.
  • llama.cpp · vLLM · Ollama · QLoRA
  • p95 latency and cost per million tokens, before / after
  • Nothing leaves your network
 llama.cpp · vLLM 03 Data & streamingPipelines that survive a restart Ingest, store, query. Kafka topics with sane partitioning, ClickHouse for the questions that used to take a night, geospatial and time-series workloads.
  • Kafka · ClickHouse · TimescaleDB · Go
  • Idempotent consumers, replayable history
  • Backfills you can actually run
 Kafka · ClickHouse 04 Infrastructure & operationsThe part nobody wants to own Proxmox hosts, containers, tunnels, DNS, certificates, monitoring, backups and the incident that follows. My own stack runs this way, publicly — you can watch it breathe.
  • Proxmox · Docker · Cloudflare · Caddy
  • Alerts that fire on a real failure, once
  • Backups restored on a schedule, not promised
 Proxmox · CF
Open to projects and to a full-time role How I work with clients
02 — Writing

Notes from systems
I actually ran.

Every number here comes from a command I can re-run. Failures included — especially the failures.

2026-08-28I Trained a 1.7B Model to Beat an 8B One — Crab-1 OSINT Agent6 min · SLM, QLoRA, OSINT, Evaluation 2026-08-05Every Improvement I Made to My 1.7B Agent Made It Worse11 min · SLM, Eval, QLoRA, Crab-1 2026-08-04I Wrote a Self-Healing Script That Could Strand the Machines It Healed11 min · Networking, DHCP, Automation, Incident 2026-08-03Two of My Own Alert Rules Were Wrong. Here Is How I Found Out.11 min · Alerting, Prometheus, Observability 2026-08-02A 2012 Mac Mini Beats a Xeon Per Core. That Changed How I Schedule Jobs.12 min · Nomad, Benchmarking, Hardware, Scheduling

All 22 notes, newest first — Writing · RSS

03 — Check it yourself

Three things that
are not a portfolio.

No screenshots, no mock-ups. Each link runs live code against real machines.

01 Live statusThe stack I run, measured from inside 30-day uptime, p95 latency, requests per day, last verified backup, open incidents. Aggregates only — no hostnames, no addresses. /status 02 Live auditPoint it at a domain, read the result A passive audit through a Netlify function with SSRF guards: DNS, TLS, headers, certificate hygiene. The same checks I run on a first call. /audit 03 Failure simulatorWhat your architecture does when a node dies A client-side model of replicas, quorum and partitions. It runs in your browser, which is also the point: no server, no upload. /simulator
04 — Contact

Tell me what you are building.

Thirty minutes, no slide deck. You describe the system — shipped, half-shipped or still on paper — and I say what I would do first, and whether I am the right person to do it. Either way you leave with a clearer picture than you came in with.

Book a 30-min call Email
Elsewhere GitHub
Network LinkedIn
Documents Résumé (PDF)
Based in Paris, Île-de-France
Yegor Gaidar · Privacy · 2026