Gaurav Upreti

Building AI systems. Writing about what actually works in production.

[Gaurav Upreti @gauptx]

@system_monarch The hardest part of “understand the plan” to actually practice.

can you explain why the AI chose approach A over the three other valid approaches it didn’t take?

If you can’t, you don’t understand the plan, you’re just reading it.

[Gaurav Upreti @gauptx]

@krunalbuilds The pattern is event-driven architecture plus a transactional outbox.

Without the outbox, a crash between saving the order and publishing the event creates a silent consistency gap.

[Gaurav Upreti @gauptx]

Graceful degradation is not one fallback.

Define tiers before an incident:

1. Full response

2. Smaller model

3. Cached/static response

4. “Try again later”

Give tiers their own SLO and telemetry.

If you cannot see which tier served a request, you cannot operate the fallback.

[Gaurav Upreti @gauptx]

@SumitM_X Same HTTP status, different machine-readable reason.

Both are 401, but token_expired tells a trusted client to refresh; token_invalid tells it to re-authenticate.

[Gaurav Upreti @gauptx]

@goyalshaliniuk LLM = model.

GenAI = capability/category.

Agent = system using models + tools + state/control loops.

ML = broader modeling discipline.

Data Science = broader data/analysis discipline.

[Gaurav Upreti @gauptx]

@konig0000 MCP → agent/model ↔ tools and context

A2A → agent ↔ agent collaboration

ACP → agent communication/interoperability, depending on the specific protocol being referenced

[Gaurav Upreti @gauptx]

@systemdesignone Nginx can reverse proxy and load balance.

Cloud/API gateways can route and balance traffic too.

So I wouldn’t classify infrastructure purely by product name.

[Gaurav Upreti @gauptx]

@system_monarch Production failures teach you what the architecture diagram leaves out.

Timeouts, partial failures, race conditions, bad assumptions, missing observability.

But the real learning is turning each incident into a system that’s harder to break the same way twice.

[Gaurav Upreti @gauptx]

@SumitM_X data, row-data, and with-config describe the response implementation.

Model the domain instead:

GET /orders/{id}

Then use query parameters or sub-resources only if configuration is genuinely optional/separate.

[Gaurav Upreti @gauptx]

@SumitM_X preventDefault() - stops the browser’s default action.

stopPropagation() - stops the event from traveling through the DOM.

You can need one, both, or neither.

[Gaurav Upreti @gauptx]

@AiCamila_ The fallback path is part of the product, not exception handling.

Different failures need different degradation policies.

[Gaurav Upreti @gauptx]

@askalphaxiv The experience representation itself becomes learnable rather than manually designed.

That could matter a lot as agent trajectories become longer and harder to summarize reliably.

[Gaurav Upreti @gauptx]

@Its_Nova1012 git blame - maps each line to the commit that last modified it.

git bisect - uses binary search between a known good and bad commit to find the regression.

[Gaurav Upreti @gauptx]

@kav_kavi11 They’re not equivalent.

salary = NULL evaluates to UNKNOWN, not TRUE, so it won’t match NULL salaries.

salary IS NULL explicitly tests for NULL.

[Gaurav Upreti @gauptx]

@thsottiaux A 1M-token window changes the ceiling, not the optimization problem.

More context can mean more latency and cost, while irrelevant context can still hurt reasoning.

The goal should be the smallest context that reliably contains what the agent needs.

[Gaurav Upreti @gauptx]

@brankopetric00 Keep the old region healthy, design both sides for overlap, and have routing/failover mechanisms planned before migration day.

The safest cutover is one where both destinations can coexist.

[Gaurav Upreti @gauptx]

@hello_code_ AI compresses implementation effort.

It doesn’t compress responsibility at the same rate.

[Gaurav Upreti @gauptx]

@DanKornas Editable memory makes debugging agent behavior much more practical.

[Gaurav Upreti @gauptx]

@konig0000 Partitioning is often a database optimization.

Sharding becomes a distributed systems problem.

[Gaurav Upreti @gauptx]

@tanujDE3180 First check whether competing transactions lock the same rows in different orders.

Consistent lock ordering is usually the first fix.

[Gaurav Upreti @gauptx]

@SumitM_X I’d choose IN, but not because it’s automatically faster.

These predicates are logically equivalent, and the optimizer will often produce the same plan.