Chocolate Milk Cult Leader| Machine Learning Engineer| Writer | AI Researcher| | Computational Math, Data Science, Software Engineering, Computer Science
New York City Metropolitan Area
15K followers
500+ connections
About
I lead the Chocolate Milk Cult — a global community of 1.5M+ builders, researchers, founders, and investors who read our deep dives on AI systems, infrastructure, and capital strategy.
Our work is read inside companies like NVIDIA, OpenAI, Google, Anthropic, and major investment firms. We focus on breaking down how AI systems actually work — the engineering constraints, research tradeoffs, and economic implications — so operators can think clearly about where the field is going.
On the building side, I’ve worked with state governments, multinational corporations, and early-stage startups. Currently working on next-generation reasoning systems, including:
Applying evolutionary algorithms inside generation loops
Exploring latent space reasoning approaches during generation
Building multi-agent systems designed for real-world production constraints.
We work on foundational algorithms and theoretical innovations to democratize intelligence, so that everyone has access to cheap, personalized intelligence.
At IQIDIS, we built the strongest legal reasoning systems available today. In live demos, we’ve gone head-to-head with heavily funded legal AI competitors and consistently outperformed them in structured reasoning tasks.
Before that, I worked across healthcare, finance, and enterprise ML — taking projects from data collection through deployment and retraining.
Some past work:
Contributed to healthcare modeling work with Johns Hopkins that informed state-level policy decisions for hundreds of millions of people.
Built automated ML pipelines that reduced manual efforts by 3-5x.
Developed high-precision customer identification systems for ICICI Bank (~95% predictive accuracy)
Co-developed a Parkinson’s detection model using voice samples under tight resource constraints (high background noise, very low resources for inference, etc)
Also, I once got 19 strangers into a tree with me. Willingly. On a Friday Night. It's really hard. Try it if you don't believe me
Skills
Machine Learning | Math| Tensorflow| Keras| Scikit| Python| Java| Docker| Git| SQL| Computer
Vision| Time Series Forecasting| Natural Language Processing| Deep Learning| Statistical Analysis| Data Science| Software Development| AGILE| Data Analysis | Big Data| Cloud Computing| Docker
Articles by Devansh
Activity
15K followers
-
Devansh Devansh shared thisMeet Cerebras, the 66 Billion Dollar IPO out to challenge NVIDIA. https://lnkd.in/ezqG7kSP An H100 executes 989 trillion floating-point operations per second. During LLM decode, the chip uses less than 1% of that capacity. The math units sit idle because the weights can't arrive from memory fast enough. This in a nutshell, is why Cerebras exists. To produce a single output token, the chip must read essentially every weight in the model from memory, multiply those weights once, and move on. The arithmetic intensity collapses to roughly 1 FLOP per byte. The H100's architecture balances compute and memory bandwidth at 295 FLOPs per byte. Anything below that ratio is bandwidth-bound, and decode is 295x below it. You could weld another GPU to the first one and decode wouldn't get faster. The obvious fix is to put the memory next to the compute. Engineers have tried this for forty years. Gene Amdahl raised $230 million in 1980 and a prototype shorted and glowed red during a demo. ETA Systems built a liquid-nitrogen-cooled wafer-scale supercomputer — the silicon worked, customers couldn't maintain the nitrogen plumbing. Tesla proved modern wafer-scale packaging works with Dojo, then Musk shut it down because the narrow workload couldn't justify maintaining a custom software stack against CUDA. Cerebras looked at all of this and decided to try anyway. They kept an entire 300mm silicon wafer intact — no dicing — stitched 84 reticle exposures into one continuous chip, embedded 44GB of SRAM next to 900,000 compute cores, solved the power delivery problem by pushing current vertically through the back of the wafer, and invented a floating connector that absorbs 150-190 microns of thermal expansion mismatch per cycle without cracking. The result: 0.168 bytes per FLOP versus the H100's 0.0034. A 50x advantage in feeding data to math units during decode. Llama 3 70B at 2,100 tokens per second versus 30-50 on a single H100. I spent about two months researching this — the semiconductor physics, the five constraints that killed every prior attempt, the first-principles bandwidth math, the OpenAI contract economics (they imply negative 81% gross margin per system on the baseline tenant), and the software ecosystem that currently supports exactly four models. Full investigation here: https://lnkd.in/ezqG7kSP Whether Cerebras survives depends on whether the market window stays open long enough for wafer-scale silicon to outrun the incumbents it forced into action. But the physics is not a debate. The incumbents already conceded the point — NVIDIA bought Groq for $20 billion, Google bifurcated its TPU line, AWS is pairing Cerebras wafers with Trainium for disaggregated inference. Someone is going to capture the value of fixing decode. The question is who. Drop your thoughts below.
-
Devansh Devansh shared thisI've been studying Cerebras for a little bit, and their engineering is so interesting. It took a few months for me to build my knowledge here, but I'm finally ready to start my deep dive.
-
Devansh Devansh shared thisAI Inference costs are killing your profit margins. Let me teach you how to reduce your Inference Overhead with Compiler & Graph Execution Running an LLM under PyTorch or TensorFlow looks simple, but the framework issues thousands of separate GPU kernel calls for every forward pass. Each kernel executes a small unit of work—like normalization or matrix multiplication—and writes the result to global GPU memory (HBM) before reading it back. While HBM bandwidth reaches 2–3 TB/s on an H100, that is 10–50x slower than the GPU’s on-chip registers. Every unnecessary trip to HBM is wasted potential. Worse, each kernel launch requires the CPU to coordinate with the GPU, adding tens of microseconds of overhead. Across thousands of tokens, this becomes milliseconds of latency. Three techniques—kernel fusion, CUDA graphs, and FlashAttention—target these bottlenecks. Kernel Fusion: Combining Operations Instead of launching separate kernels for LayerNorm and matrix multiplication, you fuse them into one. The compiler rewrites the computational graph to combine operations, ensuring intermediate results stay in the GPU’s fast on-chip registers instead of touching global HBM. This cuts memory traffic and eliminates redundant kernel launches. The tax: irregular shapes or dynamic padding can block fusion, leading to a mix of fused and unfused kernels. CUDA Graphs: Bypassing the CPU Inference involves repeating the same sequence of kernels for every generated token. Rather than the CPU re-issuing commands, CUDA graphs allow you to record the sequence once and replay it directly on the GPU. This bypasses the CPU scheduler entirely, eliminating launch overhead. The tax: graphs are tied to specific tensor shapes, requiring effective systems to capture "hot" shapes and fall back to standard execution for others. FlashAttention: Avoiding the Quadratic Wall Standard attention computes an N x N score matrix between queries and keys, which creates gigabytes of memory traffic per token. FlashAttention tiles this computation, loading small blocks of queries and keys into on-chip SRAM to compute partial attention scores incrementally. The result is mathematically identical, but the memory footprint is a fraction of the original. The tax: gains depend on sequence length, and for very short sequences, the overhead of tiling can outweigh benefits. Summary: The Performance Compound Kernel fusion ensures fewer writes and more work per cycle. CUDA graphs remove launch overhead, keeping the GPU in constant motion. FlashAttention prevents memory blowup, freeing bandwidth for compute.
-
Devansh Devansh shared thisThe memory wall is one of the most overlooked aspects of modern AI. No matter how fast we make our chips, they will always be hampered by it because of the way we produce outputs. Let's understand why. Autoregression is strictly sequential. Token 200 cannot exist until token 199 is sampled. Because of this lockstep dependency, the model cannot process its layers in parallel across the sequence. To produce a single token from a 70B-parameter model, all 140 GB of weights must stream from High-Bandwidth Memory (HBM) into the chip’s SRAM, execute a tiny matrix-vector multiplication, and leave. This operation has an arithmetic intensity of roughly 1 FLOP per byte of data moved. A modern Hopper or Blackwell tensor core needs nearly 300 FLOPs per byte before it stops starving. So when an LLM generates text for a single user, the GPU executes at under 1% of its theoretical peak compute. The other 99% of the silicon waits for data to arrive. The constraint isn’t how fast the chip can multiply. It is how fast the chip can be fed. Once you understand that data movement is the true bottleneck, the entire AI infrastructure stack reveals itself as a series of patches for this exact flaw. Take the KV cache. If you are already memory-bound, the logical fix is to avoid recomputing the past. The KV cache stores the attention states of previous tokens so each new token only adds to the sequence. But this trade creates a massive residency tax. The longer the context, the more memory capacity it eats. To offset that capacity tax, the system has to amortize the weight-streaming cost across multiple users. That is why PagedAttention and continuous batching frameworks like vLLM exist. They batch requests so the 140 GB of weights stream through the chip once to serve 64 users instead of one. Throughput scales linearly — until the KV cache for those 64 concurrent users blows out the HBM capacity entirely. When the software patches hit a wall, the hardware roadmap panics. HBM4 is a bandwidth upgrade. NVLink 5 at 1.8 TB/s per GPU is an interconnect upgrade. Multi-billion-dollar photonic interconnect acquisitions are not compute upgrades. They are data-movement upgrades designed exclusively to keep the tensor cores from starving. But the physical reality of chip manufacturing is asymmetric. Adding denser compute units and low-precision formats like FP4 is relatively cheap. Scaling memory bandwidth and capacity is economically brutal. Autoregressive inference sits on the wrong side of that divide. You can pour infinite FLOPs into a server, but if the workload requires dragging gigabytes of history across the chip for every single token, the math will never work. The only meaningul solution here is to rework inference ground up: https://lnkd.in/eAbkgrm7
-
Devansh Devansh reposted thisDevansh Devansh reposted thisWe're now at 300+ law firms and organizations on our platform. Two key learnings: First, deployment is the single most important factor that influences adoption. Clients do not get stuck on our product. It's intuitive once they see what it can do. While most firms are productive from day one, productivity and getting full value are two different things. The organizations we've closed that gap with all did one thing: they let us walk practice groups, billing, legal ops, and knowledge management through their actual workflows from day one. Adoption depth changed dramatically in the first three days, specifically: a ~30% increase in time in platform. One firm piloted everything on the market for a year before settling on us. Daily active use sits at 93%. Time in platform averages 2.3 hours. That does not happen just because the product is good. It happens because the deployment was right. Second, the people who decide whether a platform survives twelve months are not just the lawyers who signed off on it. They are operations staff, the billing team, risk and compliance, knowledge management. Operations stays for a decade. If they are not involved from day one, adoption decays quietly. A senior implementation advisor on our team deployed legal AI inside a global firm. He put it simply. Vendors do not lose because the product is bad. They lose because they never solved adoption. That pattern repeated enough that we formalized a deployment framework around it. After twelve months, the platform is not a tool our clients sparsely use. It is how their teams run. We do not install. We embed.
-
Devansh Devansh shared thisFine-tuning is expensive, heavy, and often breaks more than it fixes. Activation Steering promised a surgical alternative, but in production, it usually leaks probability into generic hallucinations. The reason? A geometric "type error." We treat AI models like flat maps, but Softmax warps that terrain. My latest deep dive breaks down the Information Geometry of Softmax and how "Dual Steering" fixes the probability leak: -)The Euclidean Illusion: Why adding vectors in the residual stream causes "probability leakage" into random prepositions. -)The Two-Map System: Understanding the Primal (freedom) and Dual (cage) coordinate systems. -)AND vs. OR Logic: Why standard steering crushes unique traits while Dual Steering preserves the union of concepts. -)The Fix: Why you can't "add a thermometer to a room"—and how to properly apply probes in dual space. Unlocking the math of knowledge representation is the path to efficient, precise control. Read the full breakdown: https://lnkd.in/eYTtVDmF
-
Devansh Devansh reposted thisDevansh Devansh reposted thisHarvey's product is well-built. But it's built for a specific environment. I know this first hand because I am one of the #legal #engineers that helped build it for over a year when Harvey first hit the market. Slaughter and May has a dedicated legal ops team, existing LexisNexis relationships, and IT infrastructure capable of handling enterprise onboarding. Their evaluation took months. Their pilots had structured criteria. The firm that can commit that kind of time and resource to a vendor evaluation is not most law firms. When Harvey wins a rigorous head-to-head at a Magic Circle firm, that result is real. It is also a signal worth reading carefully. Enterprise legal AI wins enterprise evaluations because enterprise evaluations are designed around the requirements of enterprise products. Long procurement cycles. Security reviews with dedicated IT. Legal ops bandwidth to configure and train. The product that clears those hurdles is not automatically the product that works best when a solo uploads a file at 9pm and needs a draft by morning. The failure mode I see most in mid-market adoption: firms apply Am Law evaluation criteria, then wonder why adoption stalls at 20 percent. The right question is not who won at Slaughter and May. It is whether the same architecture fits your environment. #LegalAI #LegalTech #LawFirmManagement #LegalInnovation #AIforLawyers
-
Devansh Devansh reposted thisDevansh Devansh reposted thisWhat if drafting a brief took a fraction of the time? Today, May 6 at 11:00 AM CST, we're hosting a 30-minute live session showing attorneys and legal professionals exactly how to use AI-assisted drafting for brief writing — from working with an existing case file to building an outline from scratch. In 30 minutes, you'll see how to: → Edit an existing brief with AI assistance → Draft a brief based on your case file → Create a brief outline when you're starting from zero 👉 Save your spot: https://lnkd.in/gbB7wUgb Hosted by Claire Adams
-
Devansh Devansh reposted thisDevansh Devansh reposted thisWhat if AI didn’t replace lawyers - but finally freed them to think? Devansh joined the AI for Founders podcast with Ryan Estes to explore how legal AI is shifting the profession - not by removing lawyers, but by eliminating the busywork that holds them back. The conversation dives into how AI can take over repetitive tasks, giving legal professionals more time to focus on strategy, judgment, and high-value thinking - where human expertise truly matters. Listen to the episode here: https://lnkd.in/dkjbwnvZ #AI #LegalTech #ThoughtLeadership #PodcastBooking #podcast
-
Devansh Devansh liked thisDevansh Devansh liked thisWe’re hiring builders for the U.S. team at FuriosaAI. If you’re excited about AI infrastructure, inference, semiconductors, and scaling frontier technology — we’d love to talk. Open roles: 🔹 Business Development & Sales Executive https://lnkd.in/gKATt2Bd 🔹 Solution Architect https://lnkd.in/gAEbDpCn Join us in building the next generation of AI acceleration. #Hiring #AI #Inference #Semiconductors #SolutionArchitect #BusinessDevelopment #FuriosaAI
-
Devansh Devansh liked thisDevansh Devansh liked thisLegal AI is entering a new era. We are partnering with Harvey to bring the expertise and judgment of law firms and in-house legal teams directly into AI-powered legal work. The partnership represents a step change in legal AI, combining domain-specific AI with institutional intelligence grounded in an organization’s past work, decisions, and expertise. Now, legal teams can move seamlessly from research to drafting to decision-making, informed by the full scope of their institutional knowledge. Together, every output reflects how your firm actually practices. Not a generic answer. Your answer. Full announcement in the comments.
-
Devansh Devansh liked thisDevansh Devansh liked thisThe model that impressed you last month ( especially CLAUDE) isn't the model answering you today. You're using whatever their routing layer decided your prompt was worth. An AMD executive tracked 6,852 Claude Code sessions. Measured a 67% drop in reasoning depth after February 2026. Filed a public report with receipts. The company initially closed the bug report. This gets framed as a transparency problem. It isn't. It's incentive architecture. Inference costs collapsed 200× in 16 months, $30/million tokens at GPT-4 launch, $0.15/million 16 months later. Anthropic ran at -94% gross margin in 2024. The Inference Flip happened in early 2026: for the first time, cumulative global spend on running AI officially surpassed training costs. Inference is now 85% of the enterprise AI budget. And the data shows it's load-sensitive: 5pm PST has the lowest reasoning depth of any hour. Late night recovers. Not because the model is tired, because the infrastructure is under load. Thinking allocation varies with server capacity, not with your task's actual complexity. When revenue per query craters and the GPU bill stays enormous,quality rationing becomes economically mandatory. Quantization tiers. Request routing. Dynamic reasoning depth, Anthropic's own term for a feature that reduces how hard the model thinks based on a cost signal you can't see. This is CPU thermal throttling for reasoning. Under thermal load, CPUs reduce clock frequency, same processor, reduced performance. OS monitoring tells you when throttling occurs. AI has the throttling. It doesn't have the monitoring layer. At that point you stop looking at the application layer.Pinning model versions and building evals is defensive hygiene. Not a solution. Benchmarks compound this. Frontier models score 90%+ on standard tests. Route 90% of requests to a cheaper model and benchmarks "barely move the needle." That's the phrase from the routing research. Because benchmarks measure population-level performance on fixed tasks, not per-request quality variance. At that point you stop looking at the application layer.Pinning model versions and building evals is defensive hygiene. Not a solution. Cloud computing learned the actual fix in the 2000s: SLAs with financial penalties. AI needs quality SLAs tied to measurable output, thinking depth metrics in API responses, not just latency guarantees. The data to build this already exists. The incentive to expose it doesn't. The ground remains still only for those who have proven they are willing to pay its weight in gold, (Max 20x) ;) #AIInfrastructure #MachineLearning #LLM
-
Devansh Devansh liked thisDevansh Devansh liked thisExa goes full unicorn 🦄 We have raised our Series C and are now officially a unicorn. Exa is now valued at $2.2B! This lets us keep scaling the index, push further on retrieval research, and continue growing the team. Super exciting to see what comes next! https://lnkd.in/epb_k3uw
-
Devansh Devansh liked thisDevansh Devansh liked thisReliability is one of the most important features in software, yet it’s also one of the least visible, especially in systems that are made well. On the latest episode of Alexa’s Input (AI), “Building Reliable Systems at Bloomberg with Sal Furino”, I sat down with Salvatore Furino from Bloomberg to talk about SLOs, observability, reliability engineering, and how these ideas are evolving in the age of GenAI systems. The same reliability fundamentals still apply to AI systems: - latency - availability - saturation - correctness - user expectations The challenge is that “correctness” becomes much harder to define once systems become probabilistic. Sal discusses how a trading system and an LLM assistant have very different reliability expectations, how a lot of modern engineering work is really about maintaining trust while systems become increasingly distributed and AI-driven, and more. Podcast links in comments! :)
-
Devansh Devansh liked thisDevansh Devansh liked thisInteresting take on all the hype around Mythos. I have had some of my own opinions and thoughts looking at reporting and questioned some holes I saw. But Devansh Devansh has spent some time trying to dig further down into the information and shares his thoughts. I recommend reading it, take from it what you wish. https://lnkd.in/eh9x8UQgAnthropic's Claude Mythos Launch Is Built on MisinformationAnthropic's Claude Mythos Launch Is Built on Misinformation
Experience
-
-
-
Clientell
-
-
Developer Advocate
Tech Made Simple
-
-
-
-
-
-
Rochester Institute of Technology
-
-
-
-
-
-
Education
-
Rochester Institute of Technology
BSc - Bachelor of Science Computational Mathematics with minor in Computer Science and immersion in Gender Studies
-
Licenses & Certifications
-
-
-
-
-
-
-
Honored Listee
Marquis Who's Who
Issued ExpiresCredential ID f6d768209d7a4861b10d79242315a9398128895a246e423bbd0beed294c2f3a3 -
-
-
Volunteer Experience
-
Member
The National Society of Leadership and Success
- Present 7 years 3 months
Social Services
Volunteer at various NSLS events and fundraisers.
Skills
Patents
-
Novel method for Detecting Parkinson's Disease in an Individual
Issued IN 201821005237
See patentWe invented a New Way to Detect Parkinson's disease using voice samples. Our method provides great accuracy while using lesser and lower quality data. The patent was a collaboration between myself, Dr. Aditya Abhyankar, and Suraj Patil. I am not connected to them on LinkedIn, so I couldn't add them here. I have linked the commercialization paper as proof of patent. I was responsible for the machine learning and optimization during the patent
Courses
-
Actuarial Mathematics
MATH 255 - 52675
-
Analysis of Algorithms
CSCI 261
-
Graph Theory
MATH 351
-
Introduction to Big Data
CSCI 620 - 50487
-
Linear Optimization
MATH 311 - 53371
-
Principles of Data Management
CSCI 320
-
Programming Language Concepts
CSCI 344 - 50328
Projects
-
Machine Learning Made Simple
See projectMaking Machine Learning research simple for everyone. Present actionable insights from complex theory and research in the fields of Machine Learning, Artificial Intelligence, Deep Learning, Data Science, and Programming .
-
AI: Age of Empires
-
Working on an Artificial Intelligence script for Age of Empires 2:Definitive Edition. This is helping me understand various topics such as path finding, min maxing etc.
-
Hindi Python
-
See projectDesi Python is an initiative taken by me to allow people to write Python Code in Hindi (and eventually other scripts). By allowing people to write in a script they are familiar with, we can open up coding to a large number of people.
-
Data Structures and Algorithms Visualization
-
See projectDS.dev is an initiative taken by Dan and me to make Data structures and Algorithms accessible to all. We wrote the project in React, working on various animations and works. The end goal is to allow users to create their own data structures, which we visualize. This way people will be able to see how the various data structures work
https://github.com/dl1683/DataStructuresInJavaScript -
HealthGuide: Android App
-
See projectHealthGuide is a project created for the AppJammningSummit 2017. It was created as an objective to inform people about different aspects of health. HealthGuide provided users with information about many different facets of health, allowed users to build their custom newsfeeds, had a workout recommendation system and allowed for quick queries that a user might have. This was supported by a backend supported on the Google FireBase API, and integrated with various social media APIs for…
HealthGuide is a project created for the AppJammningSummit 2017. It was created as an objective to inform people about different aspects of health. HealthGuide provided users with information about many different facets of health, allowed users to build their custom newsfeeds, had a workout recommendation system and allowed for quick queries that a user might have. This was supported by a backend supported on the Google FireBase API, and integrated with various social media APIs for authetication
Honors & Awards
-
Dean's List Spring 2021
Rochester Institute of Technology
Given to meritorious students that are able to attain a high GPA for a semester
-
Deans List: 2020 Fall
Rochester Institute of Technology
Another sem, another dean's list
-
Deans List: Fall 2018
Rochester Institute of Technology
Given to meritorious students that are able to attain a high GPA for a semester
-
Deans List: Spring 2020
Rochester Institute of Technology
Given to meritorious students that are able to attain a high GPA for a semester
-
Finalist at AppJamming Summit 2017
First Code Academy
We were part of a pan Asian Android app development competition called AppJamming Summit. Our project was selected after we won the semifinals in Singapore. The finals were held in Hong Kong.
Languages
-
French
Limited working proficiency
-
English
Native or bilingual proficiency
-
Hindi
Native or bilingual proficiency
Recommendations received
10 people have recommended Devansh
Join now to viewOther similar profiles
Explore more posts
Explore top content on LinkedIn
Find curated posts and insights for relevant topics all in one place.
View top contentOthers named Devansh Devansh
118 others named Devansh Devansh are on LinkedIn
See others named Devansh Devansh