Shubhankar Kahali
RSS Feed
I build things at the intersection of ML and people systems. I started in offensive security — reverse engineering binaries, writing exploit chains, finding zero-days across 
Now I train and deploy custom models — domain-specific transformers, graph neural networks, contrastive embedding systems — for problems where off-the-shelf models fail quietly. I also do niche headhunting, placing senior engineering and ML leaders at Fortune 500 companies. The intuition is the same in both: figuring out what will hold under real-world pressure and what will collapse the moment conditions shift.
My side project 
is where these converge: a career intelligence platform built on custom models instead of generic API wrappers. Under the hood, it combines a domain-adapted transformer pretrained on 4.2B tokens of career data, a heterogeneous graph neural network for matching across 82M+ companies and 230M+ jobs, contrastive embedding models with Matryoshka representation learning, and GRPO-aligned generative models for synthesis. The serving stack uses Eagle3 speculative decoding for 2.4x lower latency, TensorRT INT8-quantized encoders, and retraining pipelines that refresh on production feedback every 4-6 weeks. Today Glide has served 444k+ users, reached 94% match accuracy and an 81% interview success rate, and delivered $554M+ in salary value across 36.9k+ placements, with users averaging a 30% salary increase and offers closing in 16 days. Learn more about Glide →
Featured
Why Smart Companies Stop Knowing Things
Published: at 05:58 PMSmart companies rarely fail because nobody has data. They fail because reality gets softened, translated, and delayed until the organization can no longer tell what is true.
Why Neural Networks Solve the Wrong Equation
Published: at 03:45 AMPhysics-Informed Neural Networks promise to solve the equations governing reality — heat flow, fluid dynamics, structural stress. But nearly everyone enforces boundary conditions wrong, producing low training losses and physically meaningless results. Here's the fix nobody uses.
Why I Am Building Glide
Published: at 04:33 AM242 applications per opening. 0.4% success rate. Most candidates ghosted. I spent years inside recruiting watching brilliant people get crushed by a system nobody designed — so I built custom ML models, graph neural networks, and a career intelligence platform from scratch to fix it.
Where Decisions Disappear
Published: at 12:02 AMEnterprise software records outcomes. The next shift will come from systems that preserve reasoning—so organizations can remember how they decided, not just what they did.
Consciousness Is a Conversation
Published: at 03:15 AMWhat if consciousness isn't a thing you have, but a language you speak? A look at why selfhood might emerge not from computation, but from the lossy, imperfect exchange of messages between minds trying to understand each other.
Why Perfect Intelligence Must Go Blind
Published: at 09:15 AMWhy true intelligence isn’t total awareness, but the ability to leave out the right things — and how selective blindness turns overwhelming detail into decisive action.
The Momentum Trap
Published: at 01:12 PMWhy the real challenge of productivity isn't focus or discipline. It's understanding momentum. And how that understanding changes everything about how we work.
Inverse Scaling Laws in Neural Networks
Published: at 09:19 AMA comprehensive examination of inverse scaling in large language models, where increased computational power paradoxically leads to degraded performance on specific tasks, challenging our fundamental assumptions about AI progress.
Why Complex Systems Can't Be Designed
Published: at 12:08 PMWhy the biggest systems started as the smallest ones, and what Gall's Law teaches us about building things that actually work.
Teaching Machines to Think Like Machines
Published: at 02:42 PMHow RASP lets us program transformers the way they actually think, bridging the gap between neural networks and human understanding of computation.
Finding True Intelligence in Language Models
Published: at 04:54 AMWhy autoregressive language models might be more parlor trick than true intelligence, and how the search for meaningful latent representations could transform how AI understands language.
The Environmental Ceiling You Never See
Published: at 03:43 AMHow your environment silently limits your potential, and why changing your surroundings might be the most important decision you'll ever make.
The Hidden Career Advantage No One Talks About
Published: at 09:07 AMWhy the most uncompetitive career paths are the ones that require emotional discomfort, and how embracing the difficult feelings everyone else avoids can be your greatest competitive advantage.
Data's Journey to Wisdom
Published: at 11:00 AMA deep dive into how raw data transforms into actionable wisdom, and why understanding this journey is crucial for both individuals and organizations in our data-driven world.
Making AI Think Faster Without Getting Sloppy
Published: at 07:42 PMA deep dive into how we slashed AI response times using Chain-of-Thought prompting and few-shot learning, with real implementation examples and practical insights from the trenches.
Recent Posts
The People Your Process Can't See
Published: at 11:54 AMThe quiet builders who hold companies together rarely fit your performance matrix. Here's how to see, reward, and keep them before it's too late.
Bootstrapping Q
Published: at 04:54 AMA model-in-the-loop playbook for turning a low-resource language into a usable domain.
People measure your worth by their own metric
Published: at 04:34 AMWhy the way people measure your worth says more about them than you, and how understanding someone's metric for self-worth is the real key to compatibility.
When Languages Fight for Neural Territory
Published: at 01:39 PMDeep dive into dynamic mixture-of-experts for multilingual LLMs - how measuring parameter deviation reveals hidden language relationships and solves the curse of multilinguality through intelligent resource allocation.