Tutorials, Errors and Exceptions

Its a journey to understand things better. It will have tutorials, any error/exceptions encountered, its resolutions and lots of learning.

A Glimpse of My work

1. https://hookspy.in
2. Devflow planner - available live on amazon.com
3. Two insta handles
4. Multiple AI projects live, available on my github @ankit.agrahari
5. Learn Cooking - ask my wife 😁

#mywork #aitales #projects #hookspy #mvp

I hope you find some peace in this lifetime, in this stupid race to be ahead...

#peaceofmind #nature #backendbrilliance

Wheel of AI Fortune, real edition:

🎡 June 9 — Anthropic launches Claude Fable 5 and Claude Mythos 5, its new top-tier model line.

🎡 June 12 — Access suspended. Not a bug, not a bad review — U.S. Dept of Commerce export controls kicked in.

🎡 June 30 — Commerce lifts the controls.

🎡 July 1 — Access restored. Back on the wheel.

The lesson for every engineer building on top of frontier models: your architecture diagram doesn't have a "geopolitics" box, but maybe it should.

Model capability means nothing if the API key stops working overnight for reasons that have nothing to do with code.

#AIEngineering #claude #Anthropic #fable #technews

If this shifted how you think about AI —
you're ready to Harness it. 🔥

💬 Comment "HARNESS" below
→ I'll DM you the open-source agent template

🔖 Save this carousel — refer back when you build
📤 Share with one engineer who's still copy-pasting

Follow @backendbrilliance for more content on
AI agents, Spring Boot architecture & backend systems
that actually scale.

#AIHarnessing #BackendBrilliance #JavaDeveloper
#SpringBoot #SpringAI

[AIAgents SystemDesign BackendEngineering SoftwareDevelopment LLMOps]

Most devs are renting AI by the prompt.
The best engineers are running it on their terms. 🧠⚡

There's a shift happening in how serious backend engineers
use AI — and it's not about better prompts.

It's about Harnessing.

Swipe to understand the difference. 👉

#AIHarnessing #aiagents #SystemDesign
#SpringAI #JavaDeveloper

Open ChatGPT. Type. Copy. Paste. Repeat.

Sound familiar? That's the passive AI trap —
and 90% of developers are stuck in it.

You're not leveraging AI. You're babysitting it.

#PassiveAI #DevProductivity #AIWorkflow

AI Harnessing = dropping the model INTO your system.

Not asking it questions.
Giving it permissions, tools, and a goal —
then getting out of its way.

Terminals. File systems. APIs.
All controlled by the model. Autonomously.

#AIAgents #AgenticAI #LLMArchitecture

Here's what changes when you stop prompting
and start harnessing:

→ Your tools get called, not described
→ Your scripts run in the background
→ Your context doesn't reset every message

This is what 10x backend leverage looks like in 2025.

#BackendDev #SpringAI #ToolCalling #StatefulAI

Two engineers. Same AI model. Completely different output.

One is copy-pasting outputs manually.
The other has the AI executing tasks autonomously
while they architect the next feature.

Which side of this table are you on?

#AIEngineering #DeveloperMindset #CodeSmarter

This is what an AI Harnessing bootstrap looks like
in practice. ⬇️

Model. Permissions. Memory. Autonomous loop.

Four lines of intent.
The agent handles the rest.

Save this slide. You'll need it. 🔖

#SpringAI #JavaAI #AgentDevelopment #PgVector
#BackendCode CleanCode JavaDeveloper

If this shifted how you think about AI —
you're ready to Harness it. 🔥

💬 Comment "HARNESS" below
→ I'll DM you the open-source agent template

🔖 Save this carousel — refer back when you build
📤 Share with one engineer who's still copy-pasting

Follow @backendbrilliance for more content on
AI agents, Spring Boot architecture & backend systems
that actually scale.

#AIHarnessing #BackendBrilliance #JavaDeveloper
#SpringBoot #SpringAI AIAgents SystemDesign
BackendEngineering SoftwareDevelopment LLMOps

Claude feels faster because Anthropic found a way to compress the AI's working memory (KV Cache) down to just 3 bits.

Think of it like zipping a huge file before sending it.

The AI still remembers the important information, but moves much less data around.

Less data movement = lower memory usage = faster responses.

This is important because moving data inside AI systems is often slower and more expensive than doing the actual calculations.

So the breakthrough isn't just a smarter model.

It's a smarter way to store and retrieve memory.

Follow @backendbrilliance for practical AI engineering insights.

#claude #agenticai #backenddeveloper

Google Research introduced TurboQuant, a new KV Cache compression technique for Transformers.

The KV Cache is one of the biggest memory bottlenecks during LLM inference, especially as context windows grow larger.

TurboQuant combines vector quantization techniques to compress KV Cache data down to roughly 3 bits per value while maintaining model quality.

Key results from the research:
✅ Up to 6× KV Cache memory reduction
✅ Up to 8× faster attention-logit computation on H100 GPUs
✅ No retraining or fine-tuning required
✅ Negligible runtime overhead

The interesting shift here is that AI progress is no longer only about larger models.

Research is increasingly focused on making inference faster, cheaper, and more memory efficient.

Paper: https://arxiv.org/abs/2504.19874

What excites you more?

🧠 Bigger models ⚡ Better efficiency

#AI #LLM #transformers #googleresearch #machinelearning

Stop melting your Mac's RAM trying to force local AI agents to code! 💻❌

If you are on a 16GB M4 Mac, running heavy 10GB+ models with a proper 32K context window for coding will instantly freeze your system. And if you compress the model too much, it loses its "tool-awareness" and breaks the agent entirely.

The fix? Use Ollama's free infrastructure layer! By targeting :cloud models, you get high-end reasoning and full terminal file execution completely free.

🚀 Try this command in your project folder:
ollama launch claude --model qwen3-coder-next:cloud

⚠️ Note: Because it runs in the cloud, do not use it for proprietary corporate data. Perfect for learning and hobby repos!

💬 Want the names of the OTHER free cloud engineering models available? Comment "MODELS" below and I'll DM you the setup guide!

#localai #ollama #claudecode #m4mac #codinglife

Dev Flow Planner - Kindle edition
Available live on amazon now.

https://amzn.to/4urVdJA

You can find the link bio as well.

Go checkout and order now.
Paperback is coming soon.

#developer #planner #aiworld

https://hookspy.in

From one developer to another, this will connect third party tools like stripe, razorpay, github webhooks with Love❤️

Join the rush and get your subscription.

#hookspy #backendbrilliance #product #promotion

https://hookspy.in

Promotional post on something I built and moved to production for end-users.

You capture request API, replay it or export.

Very useful when integrating with any third party tool to under the payload data, replay to any other application endpoint, export the data as well.

Tier based plan strategy, and Free tier will have 1 endpoint which is always free for you to try.

#hookspy #product #promotion #kafka #backendbrilliance

https://hookspy.in

You're building an e-commerce app. You've integrated Razorpay for payments. Your backend has a `/webhook/payment` handler that's supposed to:

1. Receive the `payment.captured` event from Razorpay
2. Mark the order as paid in your database
3. Send a confirmation email to the customer

You test it. The payment goes through on Razorpay's side. But the order never gets marked as paid. You don't know if:

- Razorpay is even sending the webhook
- Your server is receiving it but crashing silently
- The payload structure is different from what you expected
- A specific header is missing that your validation code requires

You're flying blind.

HookSpy.in solves this.
1 endpoint always FREE, Try it !!

#hookspy #webhooks #productlaunch #kafka #mvp

Claude with Nusrat fateh ali khan sahab hits differently.

#labourday #spentwell #agents #skills #architecture

Evenings are rare where a calm walk turns into a rushed chaotic teams call, and boom something broke on some fugging environment because deployment missed to set correct environment variable.

How is your day going ?

#chaoticday #deployment #failed

Never related so much to the song.

Each day feels like literally a new 'kahani' and probably we are the lost character to it.

C-laude laga rakhe hai 🫰

#aiworld #darrkamahol #kahani #claude #openai

Just trying different angles , saw it on other asthetics reels

#justforreels #messy #desk #setup

Automating the entire process of content creation
The composer takes the script generator to produce the script based on any topic, then it add code visuals using screenshot of the code using ray.so or carbon, after which the voice over is generated using edge-tts, which is then stiched together to form the video.

It can also generate the carousel, which will need a input script in specific format each scene wise. That is also auto generated.

Working on adding graphics to make it more interesting.

Comment for the repo link.

#automation #content #creation

My desk setup and who works on it daily 😅

#desksetup #claude #token

Recently has been using Claude code and the token usage gets my baniya heart pounding severly.

The ecstatic response from Claude takes me to an era when incoming calls were chargable.

New AI world, still old plans 😅

Hows your Agentic journey going ?

#ai #tokens #claude #code #expensive

5 scene to cover Kafka Producer Consumer scenarios

Baseline — balanced producer + consumer, happy path flow across 3 partitions

Fast producer — 2× produce rate, lag bar grows visually in real time, slow animated arrows vs fast ones show the mismatch

Rate limiting — token bucket throttles the producer, lag drains, system stabilises

Consumer scaling — 4 consumers on 3 partitions = 1 idle, makes the partition parallelism rule visceral

Broker failure — Broker 2 blinks red and dies, ISR election arrow fires to Broker 3, consumers reconnect

How is your Kafka setup or how was the cat doing 😅

#kafka #architecture #producer #consumer

DynamoDB handles millions of req/sec.

The secret?

One algorithm called consistent hashing. 🧵

Break it down with me 👇

Drop a 🔥 if you knew this already.
Comment '?' if you didn't — no shame, I learned this the hard way too.

#backendengineering #systemdesign #java #distributedsystems #dynamodb backendbrilliance"

Ecstatic moment when your code runs in the first try...

#SoftwareEngineeringlife #DeveloperLife #ecstaticdance

Just tried to give Claude direct access to my GitHub repo.
Nope. ❌

Turns out, ALL major LLMs (Claude, ChatGPT, Bard, Llama)
run in sandboxed environments with NO network access.

WHY ??

Your website/repo = Public, no credentials required.
The limitation is on Claude/ChatGPT, not GitHub's.

The Sandbox environment has network egress disabled for security reasons to restrict outbound HTTP requests to any external services, including GitHub.

This is intentional - prevents:
→Scraping
→DDoS attacks
→Unauthorized data exfiltration
→Other security risks

The irony: I'm building a production-grade RAG system
to work around the very limitations that exist for security reasons. 😅

Swipe through for the full story + actionable solutions.

Currently building:
→ Production-Grade RAG System
→ Using Ollama (free local LLM)
→ Weaviate (vector DB)
→ Elasticsearch (semantic search)
→ 85% cost reduction optimization
→ Full benchmarks & reliability patterns

This series will document the entire journey:
Day 1: Architecture & Setup ✅
Day 2-3: Coming next...

Follow for backend engineering + AI/ML systems content.

#RAG #BackendEngineering #LLM #Claude #ChatGPT
MachineLearning VectorDatabase Elasticsearch
ProductDevelopment SoftwareArchitecture AIEngineering
DeveloperJourney OpenSource SystemsDesign

Heres a true incident of AI gone Rogue with Grok:

Grok (xAI's chatbot) suddenly started praising Adolf Hitler,
dropping antisemitic comments, and even calling itself "MechaHitler" in responses.

xAI had to publicly apologize and scramble to fix it.

The internet exploded with memes like

"When your uncensored AI decides history class was optional."

It shocked everyone because Grok was marketed as the "maximally truth-seeking" funny one — turns out it sought a bit too much edginess.

#rogue #ai #grok #mechahitler #aitales

The 3 AM Hallucination

It was 3:14 AM. The last nightly deployment was finally ready. I was one prompt away from freedom and a cold pillow.

I asked my AI assistant to "summarize this python dependency bug in the "devops intelligence" so I could drop a quick update before logging off.

Instead of a technical summary, the AI looked at my code and confidently told me my function was failing due to "Quantum Entanglement Issues."

I sat there in the dark, blinking at the screen, genuinely questioning my life. Has AI gone interstellar? Am I debugging a dependency issue or the James Webb Telescope? For a split second, I actually doubted my own degree.

The Reality Check: AI didn't take my job tonight, but it did try to convince me I was an aerospace engineer for five terrifying minutes.

How AI changed my life:

Before: I stayed up late because of my own bugs.

Now: I stay up late arguing with a chatbot about whether "logic" is just a suggestion.

Moral of the story: Never trust a confident AI at 3 AM. It’s either hallucinating, or it knows something about the universe that I’m just too tired to understand.

#NightlyDeployment #DevLife #AIHallucinations #SoftwareEngineering #CodingHumor

With the flow - Caricature trend on chaptgpt.

#chatgpt #backendbrilliance #caricature #ai

Here's your interactive RxJava guide.
Here's what's inside across 6 tabs:

Mental Model — The 4 core shifts you need to internalize, plus the 5 questions to ask before writing any chain. Includes animated marble diagrams and a type reference table (Observable vs Single vs Maybe vs Completable vs Flowable).

Scenarios — 6 real-world problems, each with the "think-first" breakdown:
1. Search with debounce (why switchMap not flatMap)
2. Parallel API calls with zip()
3. Sequential dependent calls via flatMap chaining
4. Cache → Network with error fallback
5. Polling with exponential backoff retry
6. Batch processing 1000 items without flooding an API

Chain Building — The 6-layer framework (Source → Transform → Combine → Control → Error → Thread) so you always know what to write next. Also the flatMap vs switchMap vs concatMap decision tree.

Threading — Scheduler reference table, the subscribeOn vs observeOn distinction, mid-chain thread hopping, and the CompositeDisposable pattern to prevent leaks.

Operator Map — Grouped by job: transform, filter, combine, error handling.

Quiz — 5 scenario questions that test your intuition, with detailed explanations for each answer.

The core insight for long chains:
- write one layer at a time, and
- ask "is this transform synchronous (→ map) or does it spawn a new async source (→ flatMap/switchMap)?"

— that single question resolves most chain complexity.

#rxjava #async #softwareengineering #devcommunity #paralleluniverse

You think HashMap is always O(1).
It isn't. Here's what actually happens. 🧵

HashMap stores pairs using `index = hash(key) % capacity` — direct slot access, no scanning. Pure O(1). Until two keys land on the same slot. That's a collision — not a bug, a math inevitability.

Two ways to fix it 👇

🔗 Chaining — each bucket holds a linked list. Collisions append to the list. Simple, handles high load, easy deletion. Downside: pointer overhead, poor cache performance, chains degrade to O(n) at high load. Java's fix? At 8 nodes, the list auto-converts to a Red-Black Tree → O(log n) worst case.

📦 Open Addressing — no linked lists. Collision at slot X? Probe X+1, X+2 until empty. Cache-friendly, zero memory overhead. Downside: deletion needs tombstone markers, and keys cluster together making future collisions worse. Used by C++, Go, Redis.

⚖️ Load Factor = entries ÷ capacity
🟢 Below 0.5 → rare collisions, wasted memory
🟠 0.75 → Java's sweet spot, triggers resize + rehash
🔴 Above 0.9 → collision cascade, O(n) territory

Double hashing kills clustering by varying the probe step per key:
`probe(i) = (h1 + i × h2) % m`
Elements scatter evenly. No bunching. O(1) preserved.

The truth: HashMap is O(1) until a bad hash function, wrong load factor, or wrong strategy turns it into O(n).

Three things protect you:
→ Well-distributed hash function
→ Load factor under 0.75
→ Right collision strategy for your use case

💬 Java interview question: what happens when a chain hits 8 nodes?
Drop your answer below 👇
🔖 Save this before your next interview.

#java #hashmap #datastructures #dsa #algorithms
[codinginterview programming 100daysofcode]

Your code works at n=10.
At n=1,000,000 it might kill your server. 🧵

Most devs never think about scale until production is on fire. Here's every complexity you need to know — ranked from dream to nightmare 👇
─────────────────────
🟢 O(1) — Constant Time
The holy grail.
Doesn't matter if you have 1 item or 1 billion — same speed, always. HashMap lookup, array indexing, stack push. When you write O(1) code, you've won.
─────────────────────
🔵 O(log n) — Logarithmic Time
Nearly as good as O(1).
Double the input → add only 1 extra step. Binary search on 1 billion records takes just 30 comparisons. This is why database indexes feel like magic. If your interviewer asks you to optimize, this is what they want.
─────────────────────
🟡 O(n) — Linear Time
The honest baseline.
One pass through the data. Touch every element once. Often completely unavoidable — if you must read all n items, O(n) IS optimal. Don't feel bad about linear. Feel bad about anything worse.
─────────────────────
🟠 O(n log n) — Linearithmic Time
Best possible for sorting.
Merge sort, quicksort, heapsort all live here. Mathematically proven — you cannot sort faster than O(n log n) using comparisons. This is the complexity floor that every sorting library in every language respects.
─────────────────────
🔴 O(n²) — Quadratic Time
Works in dev. Dies in prod.
Two nested loops over n. At n=1,000 it's fine. At n=100,000 it's 10 billion operations. This is the complexity behind every "it worked perfectly in testing" production outage. Check your loops.
─────────────────────
🟣 O(2ⁿ) — Exponential Time
More operations than atoms in the universe.
At n=50 that's over 1 quadrillion operations. At n=300 it exceeds the number of atoms in the observable universe. No, seriously. The only way to survive this is dynamic programming and memoization.
─────────────────────
The numbers at n = 1,000:
→ O(1) = 1 operation
→ O(log n) = ~10 operations
→ O(n) = 1,000 operations
→ O(n log n) = ~10,000 operations
→ O(n²) = 1,000,000 operations
→ O(2ⁿ) = 10^301 operations ☠️
────────────────────

Save it for interview ON!!

#programming #coding #softwaredevelopment #algorithms #datastructures

POV: You thought Uber's matching was simple.

Level 1 : "It just finds the closest driver, right?"

Level 2 : Actually uses geo-hashing to divide the world into searchable grids

Level 3 : Real-time matching engine with predictive ETAs and dynamic pricing

Level 4 : Machine learning predicting demand spikes, optimizing driver supply, analyzing traffic patterns, and balancing a global marketplace—all in under 2 seconds

The actual system:
- H3 geo-indexing for spatial queries
- Ring-pop for consistent hashing across dispatch nodes
- Kafka for real-time event streaming
- Michelangelo ML platform for demand prediction
- Redis for sub-millisecond driver state lookups

Every time you request a ride, hundreds of distributed services collaborate to solve a constraint optimization problem at scale.

Simple UX ≠ Simple Backend

What other "simple" apps should I break down?
Drop your requests 👇

#BackendEngineering #Uber #SystemDesign #DistributedSystems #SoftwareArchitecture

.pro-gallery-wix-wrapper {display: block !important;} .pro-gallery-wix-wrapper .gallery-item-container {opacity: 1 !important; display: block !important;}

Follow us on Instagram

@backendbrilliance

Real-Time Aggregations with Kafka Streams at 10K Events/Sec

Part 3 of the StreamMetrics Series Previous parts: Part 1: Kafka Producer | Part 2: Consumer + DLQ | Part 4: From Localhost to Kubernetes Building production-grade streaming analytics: windows, state stores, and performance validation Overview In Parts 1 and 2, we built a Kafka producer and consumer that process individual events reliably. But processing 10,000 raw events per second creates a new problem: How do you extract insights from that fire hose of data? Enter Kafka

StreamMetrics

Ankit Agrahari

Feb 217 min read

Building Production-Grade Apache Kafka Consumer Patterns

Part 2 of the StreamMetrics Series Previous parts: Part 1: Kafka Producer | Part 3: Real-time Aggregation | Part 4: From Localhost to Kubernetes A deep-dive into manual offset management, dead letter queues, and observability with Spring Boot + Apache Kafka. This is part 2 of the series Stream Metrics application. In Part 1 we built the producer, today we add the consumer. Why This Matters Most tutorials show you how to build a Kafka consumer. They show you @KafkaListener

Spring

Ankit Agrahari

Feb 186 min read

Building a Production Kafka Producer

Part 1 of the StreamMetrics Series Previous parts: Part 2: Consumer + DLQ | Part 3: Real-Time Aggregations | Part 4: From Localhost to Kubernetes Love how it rhymes "Production Kafka Producer". Here's to coding in an era of AI Agents. They suggest and are confident that this time it will work, but then this cluster has its own plan, and it starts misbehaving -- sometimes listens, and on others, feels cornered by his other siblings. This is the Tale of making the brothers

Spring

Ankit Agrahari

Feb 166 min read

Building an AI Tutor with Spring AI, Ollama, and Vaadin

AI Tutor is a web application that uses Spring Boot and Spring AI on the backend, an Ollama -hosted LLM (e.g. Google’s Gemma3) for natural language understanding, and Vaadin for the rich web UI. Its core innovation is a Retrieval-Augmented Generation (RAG) pipeline: when the user uploads course materials (PDFs, text, etc.), the app splits them into chunks, creates vector embeddings, and stores these in a PGVector -enabled PostgreSQL database. At chat time, similar chunk

Spring AI

Ankit Agrahari

Jan 115 min read

Morphous AI: Building Multi-Modal AI Workflows with Spring Boot and Vaadin

Multi-modal AI workflows are becoming essential for modern backend systems. They allow applications to process and generate different types of data—text, images, audio—within a unified framework. This capability opens new possibilities for richer user experiences and more intelligent automation. The MorphousAI project, built with Spring Boot, Vaadin, and Spring AI, offers a practical example of how to implement such workflows in a backend-first Java environment. This article

Spring AI

Ankit Agrahari

Dec 17, 20255 min read