Krystian Słowik

problem

An instrument pipeline took roughly 20 minutes per file before anyone could look at the contents, and results lived outside the system people actually worked in.

action

Built the chain end to end: a versioned, content-addressed ingest buffer, a native tiler integration that cut per-file processing to about 1.5 seconds, a detection service consuming tiles over RabbitMQ with a custom-trained model, results flowing into the operational app for lifecycle, billing and reporting, across on-prem GPU and cloud.

outcome

A 20-minute batch step became near-real-time, landing in the same system where the work gets managed, invoiced and reported.

what i'd do differently

CI/CD and backups first, not last.

problem

A high-value integration did not exist. Driving Plain from n8n meant hand-writing GraphQL in HTTP Request nodes and re-deriving auth, filters and pagination in every workflow.

action

Built n8n-nodes-plainapi end to end: 34 operations across threads, customers, companies, emails, notes, labels, users and CSAT, full filter and sort support, a typed credential, strict mode, zero runtime deps. Published on npm, installable from Settings, Community Nodes.

outcome

A bespoke need became a reusable node. Per-workflow GraphQL glue collapsed to one install and a credential.

what i'd do differently

Ship the node the first time I hit the gap, not the third.

problem

Every enterprise deployment re-derived the same decisions from scratch, and then hit the same incidents, which got re-diagnosed from scratch too.

action

Codified the whole path into one playbook, verified against live deployments, with every troubleshooting entry indexed to incidents that actually happened rather than ones that might.

outcome

Deployments start from the playbook instead of a blank page, and new incidents map to a known chapter more often than not.

what i'd do differently

Started it months later than I should have. The second time you explain something, write the play.

problem

At enterprise scale the easy explanation is usually wrong, and it's often already been confidently stated by someone else. Multi-month recurrences span accounts with materially different architectures.

action

Run the elimination chain, MTU, SNI, proxy env, DNS, egress paths, and read the product's source (licensing, leader election) instead of trusting the docs. Walk back the pet theory when the evidence kills it, including my own. Send the honest 'no fix, no date' update instead of overpromising.

outcome

Fixes land where the bug actually lives, and customers get evidence their own infra teams can act on.

what i'd do differently

I used to argue against the theory; now I just bring the packet capture.

problem

Agent failures look like model quality problems. Usually they aren't, but proving which layer actually broke takes reading source, not docs.

action

Trace at the failing layer: the tool-call JSON the model generated versus the schema the node expects, contexts that truncate so the agent reasons on incomplete state, provider rate-limit retries that fire wrong and cascade. Verify vendor claims at source level, including correcting confident AI-generated answers before they become commitments.

outcome

Honest diagnoses, model, prompt, or the infrastructure in between, and internal tooling so the next engineer doesn't re-derive the trace.

what i'd do differently

Trust the reproduction, not the reply. Including the AI's.