# Jiri Semmler - Product Engineer

> I'm a product builder: end-to-end and AI first. I take problems from the first customer conversation through to a working product.

Engineering Manager & Product Engineer at Keboola - I lead a backend team and still write code.

- Location: Brno, Czech Republic
- Email: jirka.semmler@gmail.com
- LinkedIn: https://www.linkedin.com/in/jirisemmler/
- Website: https://jirisemmler.eu/

I turn complex customer and platform problems into shipped product. At Keboola, I lead a backend team while contributing code across data warehouses, security and integrations. I connect customer discovery, technical decisions and rollout, and use AI in analysis, implementation and review.

## What I take on

- **End-to-end ownership.** From an unclear problem through discovery, technical design and rollout - to whether it actually works for the customer.
- **Hands-on technical leadership.** I lead a backend team and still design, write and review code. Small things I build myself; bigger ones I deliver through the team.
- **Product-engineering bridge.** I translate between customers, product and engineering, and I make the trade-offs explicit.
- **AI as leverage.** AI helps me cover more of the path in analysis, implementation and review. The decisions and the quality stay my responsibility.

## For recruiters

Short answers to the questions that usually come first.

- **Current role:** Engineering Manager & Product Engineer at Keboola. I am actively looking for a new role as a Product Engineer or Engineering Manager.
- **Roles that fit:** Product Engineer or hands-on engineering lead in a product-driven company: close to customers, close to the code, with AI as a normal part of how the team builds.
- **Where:** Based in Brno, Czech Republic. Remote-first within European time zones, with travel when it has a purpose.
- **Languages:** Czech (native), English (professional).
- **Best first step:** A short e-mail with what you are working on and where I might fit.

## Experience

### Engineering Manager & Product Engineer, Keboola (2024-present)

- Lead a backend team and work hands-on from customer discovery through implementation and rollout across authorization, data warehouses and integrations.
- Owned the Snowflake authentication migration across the platform (5k+ projects, 200k+ Snowflake users, 30+ large customers with their own accounts): planned the phased rollout, contributed backend changes and coordinated customer communication with the team.
- Introduced AI-assisted analysis, miniRFCs, multi-agent coding and automated code reviews; merged-PR throughput in the team tripled once review was semi-delegated to agents.
- Ran a "Product vibecoding" workshop for 20 engineers: how to use AI to move faster without wrecking the codebase.

### Senior Backend Engineer, Keboola (2021-2023)

- Built integrations for Snowflake, BigQuery, Teradata, Exasol, Synapse and Redshift.
- Co-designed and coordinated SOX-related controls required by Groupon: approval workflows, separation of duties and audit trails. The work unlocked a major data platform partnership.

### Lead Web Developer, Dactyl Group (2019-2020)

- Led approximately eight web developers while building and delivering client applications hands-on with PHP and JavaScript.

### Web Developer, Amadeus IT Group | Nice, France (2017-2018)

- Developed PHP/Silex and AngularJS applications in an international team, helped migrate legacy software and took over team leadership.
- The team delivered about 95% of its sprint commitments.

### Skills

- **Engineering:** PHP, Python, JavaScript, Backend APIs, Integrations
- **Data & cloud:** Snowflake, BigQuery, Docker, Terraform, AWS, Google Cloud, Azure
- **Ways of working:** Product discovery, Engineering leadership, AI-assisted development

### Education

- MSc, Information Technology, Brno University of Technology (2015-2017)
- BSc, Information Technology, Brno University of Technology (2012-2015)

## AI stack

I run a small personal AI setup across three parts of my life: home, work and my own knowledge. Each use case has its own assistant with a clear job and a point where I decide.

### From meetings to memory

Most of a working day leaves a trail: meetings, Slack threads, tickets, pull requests. Three tools turn that trail into a task list I can trust and into long-term notes. Wherever a mistake would cost something, the decision stays with me.

1. **Meetings and work signals** (capture). Meetings are transcribed locally on my laptop; no bot joins the call. The rest comes from tools I already use.
   - Meeting transcripts (Transcripted)
   - Slack and e-mail
   - Calendar
   - Linear and GitHub
2. **Scout** (work assistant). Scheduled sessions read what is new and reconcile it with what is already known.
   - Routes each transcript to its meeting and writes a summary
   - Pulls out decisions and candidate tasks; a candidate is not a commitment until confirmed
   - Keeps one running list of action items
   - Maintains a work knowledge base
   - Writes the morning briefing, meeting prep and the weekly report
3. **Sanitised harvest** (boundary). Work and private memory are separate. Only lasting lessons cross, and only as proposals.
   - Picks what stays useful: transferable experience, my own contribution, outcomes
   - Leaves out internal metrics, prices, quotes, contacts and anything under NDA
   - Clear items are filed; unclear or sensitive ones wait for me
4. **Picard** (second brain). A private Markdown vault that I and my coding agents read before starting work.
   - Sources kept unchanged
   - A maintained wiki of concepts, patterns and decisions, each linked to its source
   - Projects, goals and career notes
   - Shared skills for Claude Code and Codex

**Me: the decisions stay with a person**

- I confirm which extracted tasks are real commitments
- Unclear or sensitive notes reach me as questions; they are never filed silently
- A source that goes dark is reported as an outage, never silently skipped

**A working day, as scheduled**

- Morning: Briefing. Today's focus, open items and prep for the day's meetings.
- After standup: Standup notes. Notes and action items from the recording, plus a check of yesterday's open ones.
- Every hour: Transcript triage. New meeting transcripts are routed and summarised within the hour.
- Four times a day: Delta scans. What changed in chat, e-mail, tickets and code since the last run.
- Afternoon: Harvest and triage. Sanitised lessons are proposed to the second brain, then sorted.
- Night: Reflection. The assistant tidies its knowledge base and proposes changes to its own rules.
- Friday: Weekly report. Wins, troubles and priorities for next week.

### Other use cases

- **Code review** (Kryton; Hermes orchestration | job queue | Slack | review desk). Semi-automatic code review and approvals, orchestrated by Hermes, with my own review desk for monitoring. The team's merged-PR throughput tripled. Getting close to a dark factory.
- **Running the house** (Holly; Hermes agent | Home Assistant | Nabu voice module | Slack). Runs the household.
- **Proactive personal assistant** (Grokbot; Main day-to-day assistant). Dedicated agents running my agenda.
- **Delivery work in code** (Claude; Claude Code on my laptop). Coding agent for day-to-day delivery. It also runs the scheduled Scout sessions above.
- **Second opinion on the setup** (Codex; Codex on my laptop). A second coding agent: it audits and improves the setup itself, works with my notes and builds supporting tools.

### Tools

Warp, OpenCode, Obsidian, Scout, Transcripted, Slack, Notion, Tailscale, Linear, GitHub, Cursor, Raycast, Typeless / Wispr Flow, Home Assistant, OrbStack, Agnes

### Principles

- **Tightly scoped tools.** Each agent has one job and only the permissions that job needs.
- **Human in the loop.** Agents prepare and propose; the consequential steps stay with a person.
- **Isolation between contexts.** The work agent runs in a firewalled virtual machine with no access to my home network.
- **Monitoring from day one.** A read-only page shows whether every agent, device and review job is healthy - I watch the system instead of guessing.

## Case studies

### Snowflake key-pair migration

*When a vendor turns off passwords, it must not turn off customers.*

Keboola | 2024-2026 | technical owner

Snowflake announced the end of single-factor passwords for service accounts. For a data platform that connects to Snowflake in dozens of places, that meant rewriting authentication across the whole platform without customers' pipelines stopping. I took it on as technical owner: from the first "this is red", through design, phased rollout and my own code, to enforcement on all our accounts two years later. Along the way: one incident, one lesson about communication, and three new product capabilities instead of simply removing the old ones.

- **Scope:** 5k+ projects, 200k+ Snowflake users and 30+ large customers with their own Snowflake accounts; 14 production environments in 3 clouds, 200+ tasks across 9 teams
- **My role:** Technical owner of the programme; assignee of roughly a third of the tasks; author of the design, migration tooling and customer changelogs
- **Result:** Platform side finished and password login switched off on all accounts we operate (September 2026)

#### The problem

In autumn 2024 Snowflake published a plan to block password sign-in without a second factor. For people that means MFA; for service accounts it means a key pair or another passwordless method.

The platform connected to Snowflake on many levels: root accounts per environment, project users, workspaces for transformations and for people, components, user management in customers' own accounts, test and CI environments. Historically all of it ran on a username and password. At the same time the vendor tightened the rules for people's direct access to a shared database, which hit ways of working customers liked: SQL clients connected straight to a workspace, and BI tools connected through a writer.

No single piece was hard. What was hard: most of the business ran on Snowflake, the deadline was set by the vendor and moved several times, and any mistake meant customer pipelines stopping without warning.

#### Decisions and alternatives

In December 2024 I flagged it internally as top priority and set up a regular war room. A handful of decisions then shaped the next two years.

- **Phased rollout by risk, not big-bang.** Internal projects, then pay-as-you-go, larger customers, customers' own accounts, single-tenant - and within each phase environment by environment, by traffic. One platform-wide switch was not an option with zero tolerance for impact.
- **Two-step account migration.** Step one only adds a certificate; the password stays. Step two removes the password and switches the user type. There is a fallback between the steps, and step one could be rehearsed on every environment without risk.
- **A switch per component.** Every component got a feature flag for the new authentication. More expensive than a hard switch in code, but it let us turn things on in pieces and, above all, turn them back without a release.
- **Replace, don't remove.** When direct access to the shared database was banned, the simplest move was to take the access away. Instead three replacements shipped: a component for safely sharing data with BI tools, a SQL editor inside the platform, and a lighter dedicated-account option for customers who wanted to keep their own tools.
- **Negotiate the dates with the vendor.** Only during the incident did we learn the vendor could set an enforcement schedule per account. From then on we steered the dates ourselves.

#### My role

I was the technical owner of the programme, not only its coordinator. I wrote the design (a matrix of use case x authentication x environment type), the rollout roadmaps, and the migration commands and application migrations that switched project users on every environment. I removed the last way to register an environment with a password from the internal API, moved the platform's end-to-end tests to certificates, added key support to the user-management tooling and a bring-your-own-key option for customers with their own key policy. I wrote the customer changelogs and internal campaigns, approved releases and did the closing audit of environments.

The product and communication side was owned by a colleague from product; it would not have worked without that tandem. The rest was the team: key parts of the first phase, certificate rotation, the migration tool's UI, components and SRE provisioning.

#### Delivery with the team

- **February 2025 - phase 1.** Key pairs for root accounts, SSO into workspaces and a backfill of user types on our shared accounts. The internal reaction was sharp: late communication and unclear commercial impact. That was the first big lesson.
- **June-August 2025.** New project users are created without a password; the two-step migration of existing users runs environment by environment, single-tenants included.
- **Autumn 2025.** A public timeline, key pair as the default for workspaces, a migration tool for customers, the BI replacement component and the SQL editor in production, and the end of direct access on shared accounts.
- **Spring 2026.** Clean-up of user objects that did not meet the requirements, certificate rotation as a regular process, no more password-based workspaces.
- **May 2026 - the incident.** The vendor started enforcing earlier, and on different accounts, than we expected. In an otherwise finished migration one default parameter in a client library had not been switched. Impact: hundreds of failed jobs in about twenty projects over five hours. I found the root cause within half an hour of escalation; mitigation was switching on a prepared feature flag, with no release. I wrote the post-mortem for customers.
- **May-August 2026 - the campaign.** Fewer than a fifth of customer configurations were ready. This time communication ran in four layers: a public changelog with a how-to, a warning in the product on every affected configuration, an overview with per-customer numbers for account managers, and targeted e-mails. Adoption roughly doubled in two months.
- **September 2026 - enforcement.** Password authentication switched off on all our accounts, on three dates.

#### The result

- **Whole platform moved.** From passwords to key pairs and SSO: root, project and workspace accounts, components, user management, tests and CI.
- **Enforcement without a new incident.** Enforcement happened on all accounts we operate; customer impact was handled as ordinary support tickets.
- **Replacements, not just a ban.** Customers got replacements for the removed capabilities.
- **One incident, owned.** Root cause, mitigation and fix on the same day - and a public post-mortem.

#### What I took from it

- **Closing a migration is an audit, not an announcement.** "Done" means walking environment by environment, measuring leftover traffic on the old path and writing down the exceptions with reasons. One forgotten TODO was enough for an incident.
- **Customer communication is a product.** One line in a changelog is not enough. What worked was the combination of a how-to, a warning in the product, data for the people who talk to customers, and addressed e-mails.
- **A switch is cheaper than a rollback.** A feature flag per component paid for itself in a single day.
- **Ask the vendor.** The enforcement schedule was negotiable. Knowing that a year earlier would have saved a lot of nerves.
- **Don't migrate with a tool that needs the old way.** A migration script that signed in with a password could not migrate an environment without one. The fix was tooling that prepares and verifies the key without a live password connection.

## Contact

The best way to reach me is a short message - what you are working on and where I might fit.

- Email: jirka.semmler@gmail.com
- LinkedIn: https://www.linkedin.com/in/jirisemmler/
