Blog
Everything I have written.
How I build AI systems, what running a company taught me, and the people who actually built the technology everyone uses.
Latest · Founder stories · Sep 9, 2026The Biggest Mistake I Made as a First-Time Founder: I Was Lying to MyselfI spent five years building NeuralSpace, telling myself hustle was the price. Then a full body check-up came back: fatty liver, cholesterol off the charts.All posts 42
Ashish Vaswani's Transformer paper landed while I was sleeping in the labIn 2017 I was sleeping in the MindGarage lab to get more GPU hours for my master's thesis on making neural networks show their reasoning. Then Attention Is All You Need appeared, and my bed had to wait a few more days.
I was almost done with my thesis when the Transformer paper landedMy master's thesis asked whether a neural network could show its reasoning, not just its answer. Then Attention Is All You Need landed, and I started again.
The first AI in Practice session ran at K15t in GermanyLast month I took my AI in Practice masterclass outside my own company for the first time, to the engineering, product and design teams at K15t. Here is how it went, and what they told me afterwards.
We rebuilt our legal research agent. Lawyers preferred it 5 to 1.Jus AI 2 plans its own research, reads 75,000 documents a minute and checks its own work. An independent audit preferred it 5:1 over the version we shipped before.
You Can Only Ship as Fast as Your QA ProcessAI coding agents let our team build far faster, then QA became the constraint. Why manual QA breaks, why LLM-as-a-Judge isn't trustworthy, and how we fixed it.
We Trained Our Own Embedding Model for Legal RAG. Here's What It Actually Took.Off-the-shelf embeddings failed our legal queries, so we trained Tenet 5. Six months, one H100 node, a third fewer tokens, ~10% better quality.
I Build AI for a Living. I No Longer Believe It Can Make Real Art.I build AI for a living and used to believe it could create masterpieces. A live orchestra changed my mind. Why authentic art stays human.
4 AI Leadership Lessons from Pascal Cagni, Apple's Former Europe ChiefApple's former EMEIA chief on four AI leadership lessons: stay grounded, maximise your strengths, hire people you know, and stop calling AI a bubble.
The Reality of Agentic Coding: How We Are Turning Everyone Into a BuilderForget the hype. Here is the practical reality of agentic coding. Learn how our team uses Cursor AI to empower PMs, QAs, and legal experts to ship production code.
How My Team Managed Reduce Our LLM Inference Cost By About 50%Learn how my team team at Jus Mundi cut LLM token consumption by 46% and improved RAG quality by building a custom context aware embedding model for legal data.
I worked on this 9 years ago and today it is used in the new Meta RayBan.A look back at AirScript, my research on converting hand movements into text using EMG, and how it parallels Meta’s latest Neural Band technology for AR glasses.
Why Google's Nano Banana Pro Is The Best image GeneratorWhy Google's Nano Banana update changes the game for AI character consistency, and the 17-second workflow that keeps your subject perfect across prompts.
How I Used ChatGPT's Live Camera for Real-Time LearningTired of YouTube tutorials? How I used ChatGPT's live camera and a personalised profile to get real-time, context-specific advice while plant shopping in Paris.
I Said No to a $15 Million Acquisition Offer. Do I Regret It Now?I rejected a $15M startup offer. Five years later, I shut down the company due to burnout and AI pivots. A founder’s honest take on regret.
Google did two things this month that no one is talking about.While the world reacted to Gemini 3, Google quietly shipped Nested Learning and new TPUs. Why both matter more than the model itself.
Four Nobel Prizes for AI in one year. Who got them and why?In 2024, AI swept the Nobel Prizes. Discover how Hopfield, Hinton, Hassabis, and Jumper revolutionized Physics and Chemistry in a historic year for science.
Aakanksha Chowdhery: A lesser-known titan of the world of AIHighlights Aakanksha Chowdhery’s impact on Pathways, PaLM, and large-scale AI systems, showcasing the engineering behind today’s most powerful models.
5 Prompts That Will Help You Crack Any AI InterviewLearn how to use AI tools and smart prompts for AI interview preparation while building real depth, discipline, and hands-on mastery in artificial intelligence.
Every AI team hits a wall. Ours had ‘RAG’ written all over it.RAG looked great on paper, until we tried to scale. Here’s how Jus Mundi rebuilt its Legal AI for recall, reliability, and cost at real‑world scale.
You’re Viewing AI Wrong: Why It’s a Growth Engine, Not a Cost CutterMost companies see AI as a way to cut costs. Its real power is amplifying creativity and scale, why AI is a growth engine, not an efficiency tool.
Ashish Vaswani: The Lesser-Known Titan Who Built the Future of AIWithout Ashish Vaswani's 2017 Transformer paper there would be no ChatGPT, Claude, Gemini or Grok. The researcher who quietly started the AI revolution.
Grokipedia: The AI Encyclopedia With Zero AccountabilityElon Musk just unveiled Grokipedia, an AI-generated encyclopedia pitched as the future of knowledge. Sounds exciting, right? Until you look closer. 🤔
Fighting Imposter Syndrome - Even After Building a Successful CareerAfter shipping AI across four countries and running a startup, I still ask if I'm good enough. Rebuilding Jus Mundi's AI core, and a blind audit by Vals AI.
How I Felt While Shutting Down My Company, NeuralSpaceEver wondered what it truly feels like to close down your own startup? Here's my raw, unfiltered experience and lessons that changed me forever.
Anthropic's Model Context Protocol (MCP): I Am Not Convinced YetIs MCP the future of AI integration? What are it's pros and cons? That is MCP's promise. Is the promise of MCP being the “USB-C port” for AI real? Lets find out .
How A Three-Week AI Project with Saudi Aramco Led to Months of BurnoutFifteen-hour days on a three-week AI project for Saudi Aramco, and the burnout that followed. What the win cost, and whether it was worth it.
5 Ways to test your AI Readiness without Coding or Maths70% of leading companies struggle with AI adoption. Take the AI Readiness Test to see if you're prepared to lead, no coding or maths required.
How I Used NVIDIA GPUs to Keep Myself Warm During Peak German WintersTraining AI models warmed my room better than a blanket at -5°C. What a 2016 German AI lab taught me about GPUs, heat, and doing research on nothing.
Which One Wins? ChatGPT Search, Perplexity AI, Grok, or DeepSeek AIAI search tools are improving, but hallucinations remain. How ChatGPT Search compared with Perplexity, Grok and DeepSeek in a real-world test.
70% of Leading Companies Are Struggling with AIDiscover the real reason 70% of companies falter in AI integration: a staggering knowledge gap exposes missed opportunities worth trillions.
How I Did a Master’s in AI in Germany with Just 15K INR in Savings And No Study LoanStudying abroad without a massive loan seemed impossible, until I did it. Here’s how I earned my Master’s in AI in Germany with just ₹15K in savings.
Subtitling a million minutes a year: the ABS-CBN storyABS-CBN subtitles a million minutes a year. It took 12 hours of manual work per content hour; we got it to one, and gave them ownership of the model.
The Story of An Average Engineer - How Perseverance Can Be Your SuperpowerEver felt average in a world that celebrates excellence? How hard work and resilience turned 'average' into my greatest strength.
Why I Hired My Friends Despite Everyone Telling Me Not ToDon’t hire your friends", I ignored this advice, and it made all the difference. Through failures and success, having friends by my side kept me going.
Why I Left Canada to Build My AI Startup in OdishaI once believed AI startups could only be built in big cities, until I proved myself wrong. Here's my journey.
We told our buyers not to trust our own benchmark numbersWe published a vendor-neutral guide that undercut our own marketing: why advertised Word Error Rate means nothing, and how to test vendors on your own audio.
Why I Quit TCS (Tata Consultancy Services) in Just Four MonthsI quit TCS in just four months when they told me, "We don't care about your aspirations." That moment changed my career forever.
Building speech AI for 25 Arabic dialects when big tech ships oneArabic has 25+ spoken dialects. Almost every speech model supports one, the formal register nobody speaks. Why that gap exists and what it breaks.
The AI that reads shipping addresses in 71 languagesThe strangest thing we built at NeuralSpace: a transliteration engine for shipping labels in 71 languages, where the real job was anti-terror screening.
Our Hindi model made 2.4× fewer errors than OpenAI. Here's the recipe.Our Hindi speech-to-text averaged 18.05 WER across five public datasets: 2.4× fewer errors than OpenAI. The datasets, the numbers, and the one that mattered.
How we beat Google, OpenAI and AWS at Arabic speech recognitionA 30-person startup benchmarked its Arabic speech-to-text against eight industry giants and came first on every dataset. The method, and why they lose at Arabic.
The day Gartner called us a Cool VendorGartner named NeuralSpace a Cool Vendor in Translation Technology. What that did for enterprise sales, what earned it, and the limits of an analyst badge.