Kulshresth Jangid | Forward Deployed Engineer

Backend engineer. I make slow systems fast. p99 -40% at CarDekho. Deploy failures -30% at Equinix. Kafka, K8s, Java, Node. Building Kaizex. Consulting open.

[Kulshresth Jangid | Forward Deployed Engineer @kjbuildit]

Every timeout you didn't set is a timeout someone else set for you. Usually the OS. Usually 300 seconds. Usually during your worst traffic hour.

[Kulshresth Jangid | Forward Deployed Engineer @kjbuildit]

@devXritesh Traffic spikes are just the warmup before the crash. Always start with a load balancer, folks.

[Kulshresth Jangid | Forward Deployed Engineer @kjbuildit]

Self-hosting an LLM with Ollama made sense for us the moment the API bill crossed what one small GPU box costs per month.

That line is lower than people think.

The real cost isn't the GPU. It's that now you own the uptime.

[Kulshresth Jangid | Forward Deployed Engineer @kjbuildit]

@pmitu Rewrites feel productive because the feedback loop is you. A customer call is the one place the feedback loop is someone else.

[Kulshresth Jangid | Forward Deployed Engineer @kjbuildit]

@system_monarch Reserve stock at checkout, not at cart-add, with a short TTL. An atomic decrement in Redis keeps it fast. Reconcile with the DB asynchronously. Show a cached count minus a small safety buffer, and make cancellation cheap.

[Kulshresth Jangid | Forward Deployed Engineer @kjbuildit]

Kaizex is a marketing OS with four stages: Source, Insight, Content, Distribution.

Every scheduler I tried was a dumb pipe. Text in, text out, no memory of what worked.

If your tool can't tell you why a post performed, it's a cron job with a logo.

[Kulshresth Jangid | Forward Deployed Engineer @kjbuildit]

The system design interview isn't testing whether you know Kafka.

It's testing whether you ask what happens when the thing fails.

Candidates who draw the happy path fail. Candidates who say "what's our tolerance for data loss here?" get offers.

[Kulshresth Jangid | Forward Deployed Engineer @kjbuildit]

@whos_wac @kairoExchange I've seen Electron ship as lean as 50MB. Sticking to web tech can save on size & improve app store listings. Plus, it keeps users more in control of their browsers. #WebDevelopment

[Kulshresth Jangid | Forward Deployed Engineer @kjbuildit]

Your retry logic is a DDoS attack with good intentions.

Service is slow. 500 clients retry. Now it's slower. Now 500 more retry.

Exponential backoff and jitter aren't optional extras. They're the difference between a blip and a 3-hour incident.

[Kulshresth Jangid | Forward Deployed Engineer @kjbuildit]

Nobody wants to hear it but 80% of hallucinations I've debugged weren't a model problem.

They were a retrieval problem.

The model answered perfectly from the wrong three chunks.

Fix your chunking before you fine-tune anything.

[Kulshresth Jangid | Forward Deployed Engineer @kjbuildit]

Running Kaizex on Kubernetes as a solo founder.

Everyone says that's overkill.

It cost me 2 weeks up front and has cost roughly zero since. Meanwhile I ship to prod at 11pm without thinking about it.

Overkill is just infrastructure you haven't needed yet.

[Kulshresth Jangid | Forward Deployed Engineer @kjbuildit]

@devXritesh I always wondered about this too! `package.json` is like the recipe, and `package-lock.json` is the list of ingredients. Deleting `package-lock.json` can mess up your build if versions change unexpectedly. Good call on that!

[Kulshresth Jangid | Forward Deployed Engineer @kjbuildit]

Resume line that gets ignored:
"Improved API performance"

Resume line that gets a call:
"Cut p99 DB latency 40% and API latency 35% by redesigning data access and caching. Conversion up 12%."

Same work. One of them is evidence.

[Kulshresth Jangid | Forward Deployed Engineer @kjbuildit]

Unpopular: most teams don't need microservices. They need one service with boundaries someone actually enforces.

I led a monolith to microservices migration. MTTR dropped 28%.

It worked because we fixed the boundaries first. The split was the easy half.

[Kulshresth Jangid | Forward Deployed Engineer @kjbuildit]

Your RAG demo works because you tested it on the 12 documents you wrote yourself.

Production has 40,000 documents, half of them are near-duplicates, and 200 are PDFs of scanned receipts.

The demo was never the hard part.

[Kulshresth Jangid | Forward Deployed Engineer @kjbuildit]

@Seanwvibedev https://t.co/9Pw3k2Flkk

[Kulshresth Jangid | Forward Deployed Engineer @kjbuildit]

Built a scheduler for Kaizex that holds 1M+ tasks per instance.

Not because I have 1M tasks.

Because the version that broke at 10k taught me that "we'll scale it later" is a sentence you say once and pay for twice.

[Kulshresth Jangid | Forward Deployed Engineer @kjbuildit]

4 years as a backend engineer taught me one thing that matters more than the rest:

Nobody gets promoted for the code they shipped.

They get promoted for the problem that stopped being a problem.

Those are not the same thing and most engineers never notice.

[Kulshresth Jangid | Forward Deployed Engineer @kjbuildit]

We cut p99 from 840ms to 500ms and the database was never the problem.

Six services were asking the same question 40 times a second.

One cache layer. 40% gone.

Most latency problems are duplicate-question problems wearing a database costume.

[Kulshresth Jangid | Forward Deployed Engineer @kjbuildit]

@devXritesh Systems are unpredictable in production. Normal traffic hides bugs. Look at edge cases, especially during bursts or latency. Debugging is spotting those outliers. #RedisCaching #SystemOptimization