SoftMax An engineering studio. Two pillars: complex real-time backend systems — experience at ~20M requests per second at 10–80 ms latency — and AI built into your product: chat, knowledge retrieval and agents on official SDKs, with cost limits and measured quality.
// facts from the CV — verifiable
Two pillars. First: complex backend systems — distributed, low-latency, running 24/7. Second: AI built into the product — chat, knowledge retrieval, MCP servers and agent systems on official provider SDKs, under cost limits and with measured quality.
Pillar 01 · Complex backend systems
Real-time services under real load — designed for scale.
// scale from practice, not a slide — billions of auctions daily in AdTech.
A backend you can grow: clear service boundaries and deliberate data.
// architecture and scalability are the day job, not a brochure line.
Systems that survive traffic — and can be diagnosed.
// reliability, observability and production debugging — the areas we work in daily.
Pillar 02 · AI in your product
An assistant that knows your product — from prototype to production with cost limits.
// model cost and response time get a budget — just like every request in the backend.
Answers grounded in your documents — with the source cited, not the model's memory.
// no source = "I don't know". Retrieval quality is measured on control questions.
Process automation that doesn't guess — behind contracts and quality gates.
// a prompt change goes through regression tests like any code change.
Every line: custom work after a consultation. Discovery call — 30 minutes, free.
Write to usHow we design
In real-time systems a late answer is worth as much as no answer. So we divide the latency budget into stages up front — and hold it by measuring the tail of the distribution, not the average.
FIG. 01 · LATENCY BUDGET
example budget · illustrative numbers
↳ off the critical path: event write to Kafka (asynchronous)
FIG. 02 · THE SAME DISCIPLINE IN AI
Token cost and model response time get a limit too — nested.
Going over = a clear message for the user, not a surprise invoice at the end of the month.
Every stage has its own limit. An overrun shows up in the metrics at once, not in customer tickets.
Event writes go asynchronous, reads come from cache (Redis, Aerospike).
Past the deadline we answer with a default variant. Degrade, never fail.
p99 and p99.9 from production — not an average from a laptop benchmark.
// ~20M requests/s · 10–80 ms · 24/7 — the scale this discipline comes from.
How we work
We don't start with throwaway MVPs. We start with what's going to production — so the quality is there in version one, not version five.
More test code than production code. BDD for user flows, real external APIs in tests over mocks wherever possible. No tests, no deploy.
You get the repository, the docs and a record of architecture decisions — not a closed box. Anyone can do the code review.
GDPR, KMS for sensitive data, immutable audit trails, versioned ToS. Ready for any regulator audit at any moment.
ShedLock crons, Redis sliding windows, dual-layer rate limiting, optimistic + pessimistic locking. Scalability isn't an add-on.
Immutable deploys, SHA-256 verification, auto-rollback on health-check failure. No attack window on production.
The tools we work in daily. Always chosen for the task — never the other way round.
Because small = lower margin on their infrastructure: project managers, stand-ups, offices, account managers. We have no infrastructure. We have code.
Code review, security check, performance audit. Concrete document at the end.
Proof of concept, integration, MVP feature. Working code in two weeks.
Discovery → fixed scope, fixed price → delivery. No T&M creep.
You pick the scope. We tell you what it costs.
Write to us — intro call (30 min, free) →Every project we ship to production is regulator-audit-ready. Not retrofitted. From commit one.
Versioned consent chain with versioned ToS, withdrawal audit in a single SQL query. Aligned with UODO 2024 guidance.
GDPR Art. 17. Sensitive data KMS-encrypted — deleting the key revokes access. 30-day execution.
Every ToS version pinned to the user account. Dispute = show consent on day X. No retroactivity.
Fakturownia → KSeF ready from the mandate start (2026). Idempotent queue, exponential backoff retries.
Regulator audit? 24h to prepare. Immutable, signed audit trail kept for 6 years.
Distributed Systems / Software Engineer. 5+ years at RTB House — systems processing ~20M bid requests per second at 10–80 ms latency, 24/7. Previously Microsoft (Prague, Teams — the follow-up-meeting feature), 9livesdata, Samsung. Stack: Java, Kafka, Python, GCP, BigQuery, PostgreSQL, Aerospike. Three University of Warsaw degrees: MSc Computer Science and JSIM — a selective MIMUW track open only to olympiad finalists and laureates (BSc Math + BSc CS in 4 years, GPA 4.63/5 with honors). Bronze Medal at IPhO 2014, Laureate of the Polish Mathematics Olympiad 2015 (16th in Poland), Finalist of the Polish Informatics Olympiad 2015.
„I don't build things I wouldn't run in an environment where everything breaks."
An intro call — 30 minutes, free. We arrange it over e-mail: write a few sentences about the project and we reply within 24 business hours with proposed times.
To: softmax.contact@gmail.com Subject: Intro call — [company name] Hello, we would like to talk about: [e.g. a backend system / an AI chat in our app / a code audit] A few sentences about the need: [...] Convenient times: [e.g. Tue–Thu, 10:00–14:00] [name, company]