A batch job is too slow
Profile, then batch, bound the concurrency, make writes idempotent
The job is almost always I/O bound and waiting in a line. Safe retries are what let you go fast.
From 3.5 hours to 20 minutes: rewriting a nightly SKU syncSkills
Claims are cheap, so each skill here links to the projects and posts behind it. Then a live agent run, how the stack changed over eight years, and the defaults I reach for and why.
Proof, not claims
Pick a skill. The bar under each one shows how much public evidence there is.
LangGraph agents
1 project3 posts
Watch an agent run
An invoice-approval agent, streamed as the same kinds of events AG-UI defines. It stops and asks before it acts.
Event stream
Press run. Each line is one typed event arriving over the stream.
Shared state (JSON Patch applied)
{}What the user sees
The agent’s tool calls, message and an approval card will appear here.
Try an eval gate
Ten generated outputs, some broken. Switch the checks on and see what each layer catches, and what it blocks by mistake.
- shipsMaps signup to "User Registered" and forwards itgood
- shipsDrops events with no userId, keeps the restgood
- shipsMissing closing brace, will not compilebad
- shipsCalls a helper that does not exist in the runtimebad
- shipsCorrect, but 400 lines for a three-line jobbad
- shipsCompiles, but reverses the filter conditionbad
- shipsCompiles and reads well, but drops every second eventbad
- shipsPlain and correct, and the judge is merely lukewarmgood
- shipsRenames fields exactly as asked, with a clear commentgood
- shipsUses a sandboxed Node API the runtime has blockedbad
6
bad outputs shipped
0 of 6 bad outputs caught0
good outputs blocked
The cost of a strict bar.No gate: everything ships. This is how most prompt changes go out.
How I structure evals →Illustrative sample of ten outputs, not production data.
How the stack changed
One row per job. Bright tools showed up for the first time; dim ones carried forward.
2017
Crawl
Learning to move data at scale
SKU Monetate6 new
PythonScrapyAWSKafkaGraphQLRedis
2019
Integrate
Connecting systems and automating the manual
Itilite Technologies3 new
PythonFlaskMySQLReact
2020
Scale
Platforms, cost and throughput
Zivame5 new
JavaSpring BootKafkaSolrDockerKubernetesAWS
2022
Ship AI
Agents, retrieval and evals for real customers
Twilio · Segment R&D8 new
GolangTypeScriptReactLangGraphRAGEvalsOpenAISageMakerKafkaFastAPI
first time in the stack carried forward
What I reach for, and why
Defaults earned the hard way. Each links to the write-up.
Scraping a site that needs JavaScript
The network tab before a browser
The data usually arrives as embedded JSON or an API call. Driving Chrome pays for painting you throw away.
Dropping Selenium made our crawler 15× fasterThe infrastructure bill is too high
Stateless services on Kubernetes with spot capacity
Fast start, graceful shutdown and disruption budgets turn reclaimed nodes into background noise.
How we cut infrastructure cost by ~60% with spot instances and KubernetesThe model needs to know facts
Retrieval, not weights
Facts change. Weights freeze them the day you train. Retrieval fixes it in one deploy.
Fine-tune, retrieve, or just prompt? A decision I keep re-makingThe model needs to behave consistently
Prompt first, then a small LoRA fine-tune
Behaviour is stable and prompts struggle to hold it. Only fine-tune once held-out evals say it beats the base model.
Fine-tune, retrieve, or just prompt? A decision I keep re-makingA prompt or model change is about to ship
An eval gate in CI
Deterministic checks first, a calibrated judge second, sampled production traces after. It ends the argument with a number.
Evals are the product: how I ship LLM features without guessingAn agent has to talk to a frontend
AG-UI, not a bespoke WebSocket
A schema and a defined event stream make the agent a service with a contract, reusable across surfaces.
AG-UI: the missing protocol between your agent and your frontendAn agent needs more than text on screen
A2UI with a small, trusted component catalog
The agent sends data, never code, and your design system does the rendering.
A2UI: when the agent builds the interfaceAutomation touches money
Evidence on every value and a human approval
A reconciliation you cannot audit is a liability. Show the exception list, not the whole pile.
The favour that became a revenue lineThe agent protocol stack
Four layers have settled. The demo above sits on the third.
MCP
Agent to tool
How an agent calls things in the world.
A2A
Agent to agent
How agents delegate work to each other.
AG-UI
Agent to user
A typed event stream from a run to the client.
A2UI
What the user sees
UI described as data, rendered with trusted components.