Deploying ML models to production
From Jupyter notebooks to scalable inference endpoints on AWS SageMaker and Lambda — a production-ready deployment playbook.
Tech Blog
Field notes from the fleet
Deep-dives and hard-won lessons from the SOSW engineering team — serverless on AWS, event-driven data, IaC, ML in production, and the spec-first systems our agents build and run.
From the blog
Notes, since the beginning
Where we started and where we're going — reposted from our channels and dated as first published.
Agents
Spec-first development with an agent fleet
Most of the code we ship is written by agents, and the reason we can trust it has nothing to do with which model wrote it. It is the spec that comes first.
Read →Agents
The harness matters more than the model
Autonomy is a dial, not a switch. Field notes on the three operating modes we run AI agents in at SOSW — and why the permissions, review gates, and budgets around an agent matter more than which model is inside it.
Read →AWS
One SAM template, many AWS accounts
Dev and prod live in separate AWS accounts, and the same template must deploy identically to both. The rules that make that possible — no Stage in names, no default Stage, exports as contracts — are the same rules that make templates safe for AI agents to write.
Read →Architecture
When event-driven architecture makes sense
Event-driven architecture buys you decoupling and free retry semantics, and charges you in debugging pain and consistency surprises. Field notes on when the trade is worth it on AWS, and when a boring synchronous API is the right answer.
Read →Big Data
DynamoDB patterns for high-throughput workloads
Field notes from running DynamoDB at sustained high throughput: partition key design, single-table trade-offs, GSI costs, write sharding, capacity economics, and the failure modes nobody puts on the architecture diagram.
Read →Team
From student to engineer: congratulations, Yulia
Yulia Rozova finished our six-month course — and joined the SOSW development team, working on real projects.
Read →Training
Why SOSW?
Where SOSW started: teaching the cloud skills the market actually demands — small cohorts, real partner projects, and a path into the industry.
Read →Long reads
On Medium
Long-form articles by our founder, Nikolay Grishchenko — from the original sosw orchestration ideas to how AI changes the engineering career itself.
Careers
Junior developers must evolve: advanced programming without writing code
For many years, the career path for a junior developer was clear and predictable: learn a language, write a lot of code, fix bugs. That path is gone.
Read on Medium ↗Security
Authentication vs authorization
Though these terms may sound very similar, there is a significant difference that should not be ignored when designing systems.
Read on Medium ↗AWS
sosw — AWS Lambda orchestration: Twitter API tutorial
The first in a series exploring "uncommon" use-cases of serverless applications — orchestrating Lambda workers with the sosw framework.
Read on Medium ↗Architecture
Serverless orchestrator of serverless workers in AWS
How we solved managing serverless workers in AWS with both the workers and the orchestrator running as Lambda — the idea that became sosw.
Read on Medium ↗The backlog
Topics we cover
The ground we work on every day. Each becomes a post when there's something worth saying that you won't read elsewhere.
Cloud security fundamentals
IAM, VPC, encryption at rest and in transit — the non-negotiables when building production systems on AWS.
Zero-downtime deployments on AWS
How to set up blue/green deployments using CodeDeploy and API Gateway canary releases for serverless applications.
What an agent audit trail must record
If an agent ships code, you need to answer who decided what, when, and why — the event log that makes autonomy reviewable.
Choosing models per role in a fleet
Frontier models for design and review, cheap fast models for the grind — how we match LLMs to agent roles and budgets.
Onboarding a team to agent operations
The augmented mindset in practice: what developers, reviewers, and managers each need to learn before the fleet pays off.
No queue, no waiting
Want the deep-dive on your stack instead?
We'd rather build it with you than write about it. Bring one workflow and we'll spec a pilot on your real data.
Start a project →