I build AI and AI-powered products. My focus areas are healthcare, making SMBs operationally effective and algorithmic trading.

Bangalore
Joined March 2009
Bargava retweeted
We’ve just made public 125,091 DICOM radiology images from across my cancer journey (link below). The data is both longitudinal and multi-modal, dated and linked to my treatment timeline, so you can compare imaging across initial diagnosis, standard of care, first remission, “founder mode cancer care”, and second remission. It’s viewable in OHIF, the standard open-source viewer used in radiology research. Given the long history of applying deep learning to radiology, we’re hoping to see a variety of methods and models applied to my data. If you find something interesting, please reach out to [email protected], or if you know a researcher who should see this, tag them below. Thanks to Terra LaFerriere for organizing these images, and Kamil Slowikowski for building osteosarc.com into the resource it is.
19
74
8
504
28,415
What timing! We built and released something similar (perhaps better) last week. We call it the Latentforce Workspace - workspace.latentstack.dev
Now in beta: composer predictions in Codex for Pro users. Codex can now suggest your next message based on your conversation and how you talk to it. One of the most loved new features we've ever tested internally.
4
12
88
7,774
It is important for me to share why I'm continuing to build Pravalclaw, in this context. I think an openclaw like assistant is still a better bet, despite the better harness engineering that seems to be the case in Muse and Grok. Something which allows user control and greater security / compartmentalization of user data will any day be better. My experience building Pravalclaw, a product I have come back to after a few months, is a good example of this. Building the agent after the initial scaffolding so that it can adjust to your workflows is a very gratifying process and high leverage. Developers who miss out on this process of customizing the agent to their needs (even if it is a bit slower than just adopting one of the existing agents) is a better long term strategy. Open model deployed on local, harness engineered by yourself or managed by yourself locally, and an interface through chat over Signal or a similar interface is a better idea than just giving into the likes of Muse and Grok Bot in my opinion. Pics attached - 1. Pravalclaw helping me summarize a document and add it to my personal knowledge base. 2. Pravalclaw deploying a website on local from a prompt. 3. Deployed website created by Pravalclaw. cc: @Fintech03 @meerasndr
The personal agents - like Muse, Instinct etc are a bigger threat to large consumer companies than the underlying AI itself. Rent seeking will be impacted in the short term. Good post from @chamath
3
1
1
5
391
Need help 👋 I'm doing a small study to find out progress in LLMs ability to write original jokes. Please participate in the study below 👇 Will take ~10 mins. I'll email results to all participants first. --- Context: LLMs are making insane progress in verifiable domains like math and coding. But what about non-verifiable domains? I took humor as an example of it and want to find out if frontier models like Astra are able to produce better original high-quality jokes v/s older models. LLMs tend to memorize jokes on the Internet, so I forced novelty by asking models to always include two randomly chosen words in the generated joke. I have LLM-as-judge results for generated jokes, and they show a clear trend of progress over time. However, I need human raters to ground the judgements. So please participate in the study (link below). You might get a chuckle or two.
17
4
54
9,957
Bargava retweeted
thanks to everyone who joined my session at @aiDotEngineer paris yesterday! for everyone who missed it, here's the thread form of my talk 🧵
7
6
3
83
6,640
Bargava retweeted
“Jev can chat too”
87
289
47
5,102
200,057
Introducing Bespoke Nimble: an open data, open model, open recipe for an open Jev. Code and info: github.com/bespokelabsai/nim… Model: huggingface.co/bespokelabs/B… Data: * A new data curation recipe called contrastive data curation. * Slightly change facts to generate negative data. This pushes the model to discriminate better and become a better decision maker. The calibration is implicit. * Didn't do ablations but I think this is a critical piece! * This also means training data doesn't need probabilities. * Data covered 10 categories, and is fully synthetic. * This data is split into train and eval. Training * LoRA finetune of Qwen3.5-9B. * Distillation-free: we use Jev to only evaluate. * No RL yet! Serving * Parallel constrained decoding as suggested by @NielsRogge and @harshagundal. Results: * The post-trained Qwen (Nimble) became substantially better on our curated eval: 66% for Qwen to 90% for Nimble. Jev is at 93%. * 100ms on H100 and free to use on your macbook! Feel the AGI for free. * 2 days of building in public. :) Big caveat is that there is no standard benchmark to measure performance, and it's possible Nimble is much worse on other benchmarks compared to Jev. But it should be better than Qwen! We thank @typesafeai for making Jev and the inspiring discussions in the community. Hope this release lifts all the boats and encourages more research and activity in this space.
Should I build an OpenJev?
82
177
40
1,394
556,905
Don't know what Navier-Stokes Equation are? It's basically Newton's law of motion applied to small parts of fluid - liquids or gases, which might be compressible, viscous.
1
7
23
1,746
I just published my complete guide to reinforcement learning for LLMs. It's a single, standalone resource for understanding RL from first principles to frontier research. Read it here: cameronrwolfe.substack.com/p… The topics covered are as follows… RL Fundamentals: - A general framework for RL (policy, states, actions, transition functions, environment, rewards, trajectories, etc.) - Core concepts of RL (returns, discounting, action / state / trajectory probabilities, value / advantage functions). - Formulating and estimating the learning objective for RL. Basic of RL for LLMs: - Token-level (MDP) vs. completion-level (bandit) formulations. - Outcome versus process rewards. - Value estimation via value models / critics. - KL divergence, RL setups (RLHF, RLVR, infra, etc.), importance sampling. Basic Policy Gradients: - Deriving the RL objective / vanilla policy gradient (VPG) from first principles. - Different policy-gradient formulations (trajectory returns, reward-to-go, baselines, value functions, and advantages). - Why policy gradient estimates have high variance (and how baselines help). - Implementation of VPG. REINFORCE and its variants: - REINFORCE as a Monte Carlo implementation of the VPG. - Completion-level vs. token-level implementations of REINFORCE. - Reward baselines and KL-regularized rewards. - RLOO, REINFORCE++ and other critic-free policy gradient algorithms. Actor-Critic Methods: - Training value models / critics alongside the policy to estimate advantages. - Trust Region Policy Optimization (TRPO) and KL-constrained policy updates. - How PPO simplifies TRPO with a clipped surrogate objective. - Full PPO details (clipping, advantage estimations, value loss, KL, policy ratios, multi-epoch optimization, etc.) + implementation. - Generalized Advantage Estimation (GAE) + implementation. - The memory and computational overhead of actor-critic algorithms. GRPO: - Group Relative Policy Optimization (GRPO) as a simpler alternative to PPO. - Group-relative advantage estimates + implementation. - Why removing the critic is helpful. - Limitations of vanilla GRPO algorithm. - Dr. GRPO, DAPO, GSPO, and other recent improvements. Recent research topics: - Online versus offline RL. - Rubric-based RL. - RL scaling laws. - Agentic RL + world modeling. This post is a synthesis of my writing / learning on RL for over a year. I hope it’s helpful!
40
205
14
1,346
62,093
Bargava retweeted
My cancer case is n=1. I'm not holding it up as a model, because most of what I did is not specifically replicable or scaleable. But it is a proof of concept that fully personalized medicine, built around one patient's own biology, is viable in ways it simply wasn't a few years ago. I have lived a version of the future of cancer care that I think is worth aspiring to more broadly. Many of the tools that made this future possible for me are getting better, faster, cheaper, and more accessible every year. I think what we need next is the infrastructure to deploy these tools, and the community willing to build this infrastructure, one patient at a time. I want to help that future arrive faster. I wrote an article in The Cancer Letter cancerletter.com/guest-edito… for the physicians and researchers closest to the patients who could benefit most from this approach, who also believe in this future. If you are working towards building the systems and practices that will make personalized medicine operationally real, we want to collaborate with you too.
49
150
18
1,043
225,614
Bargava retweeted
I see lot of you're using this phrase "software factory" here, but I don't think the phrase means what you think it means
1
4
964
🚀 Praval 0.8.1 is officially launched. Completely rebuilt the model execution path to give Praval AI agents freedom from provider specific logic 🔥 What’s new? ✅ Universal ModelRuntime: Seamlessly switch between OpenAI, Gemini, Anthropic, Cohere, & local servers. ✅ Native MCP Support: Discover and route tools effortlessly. 🛠️ ✅ Multimodal & Voice: Native support for images, files, audio, and video inputs. 🎙️ ✅ Rock-solid HITL: Built-in Human-in-the-Loop approvals give your models room to think without giving them unapproved authority. 🛑 Also added 13 course notebooks + 4 capstone projects to get you building production-ready multi-agent systems today. 💻 Check out the full deep-dive on the architecture changes here 👇 Read the full Praval 0.8.1 update pravalagents.com/blog/index.…
1
1
4
143
Looking forward to Bangalore Paper Club - Edition #2, co-hosted with Conscious Engines (@c_engines), this Friday. This week, we're diving into the following papers: •⁠ ⁠LLaDA: Large Language Diffusion Models •⁠ ⁠Nemotron-TwoTower: An Efficient Architecture for Diffusion Language Models •⁠ ⁠A Graph Talks, But Who's Listening? Rethinking Evaluations for Graph-Language Models •⁠ ⁠Canine Olfaction Combined With Bayesian Modeling for Multicancer Detection From Breath Samples: A Phase II Study in India Together, we'll go beyond the abstracts to unpack the derivations, architectural decisions, experimental methodology, and what actually holds up in practice. We're almost at capacity, with just a few spots remaining. RSVP here: luma.com/xwono22w @kautukkundan @SeekingN0rth @nityabaskar
1
5
2
56
25,252
I taught JEE physics for years. That paper breaks strong kids in three hours. This exam is five hours of theory and five hours of lab work, and these five did close to perfect scores on it. Let me tell you what actually happened. The International Physics Olympiad is the world championship of school physics. It was the 56th edition. Held in Bucaramanga, Colombia, from July 5 to 12. 381 students. More than 85 countries. Every one of them the best physics student their country could find. India sent five kids. All five came back with gold. Their names are Kanishk Jain from Pune. Riddhesh Anant Bendale from Indore. Rishit Garg from Dwarka in Delhi. Shresth Suraiya from Mumbai. Svarit Joshi from Ahmedabad. We know a hundred cricketers by their nickname and not one of these boys. :) That clean sweep put India at joint World Number One. Tied with China, Russia, Kazakhstan, South Korea and Taiwan. Those are countries that pour serious money and national pride into science education. We are standing level with them. Now here is what the exam actually was. Two papers. Each five hours long. The theory paper had three problems. One on the thermodynamics of paramagnetic cooling. One on the photoionisation of ozone. One on the dynamics of electron positron pairs. The experimental paper was another five hours in a lab, working through heat transfer and thermodynamic processes in fluids. That means you get given equipment you have never seen, and you have to design your own experiment, take your own readings, handle the errors, and reach a real answer. Not multiple choice. No shortcuts. No pattern recognition. You either understand physics or you sit there for five hours. HBCSE says the Indian students were near perfect on theory and excellent on the practical too. Now, this was India's 27th appearance at the IPhO. Across all those years, about 44 percent of Indian students have won gold, 41 percent silver, 10 percent bronze. In the last ten years, every single Indian student has come home with a medal. 62 percent gold, 38 percent silver. Not one kid has gone and come back empty handed in a decade. Five golds in one year has happened only twice. This year, and in 2018. So who built this. The programme is run by HBCSE, the Homi Bhabha Centre for Science Education. It sits under TIFR, which sits under the Department of Atomic Energy. They run the whole funnel. A national exam, then a national olympiad, then a brutal selection and training camp, and out of everyone in the country, five kids get on a plane. The team was led by Professor Anwesh Mazumdar of HBCSE-TIFR and Dr Leena Joshi from St Xavier's College, Mumbai. The scientific observers were Professor Ananda Dasgupta from IISER Kolkata and Nisha Kelkar from Gogate-Joglekar College in Ratnagiri. Yes. Ratnagiri. A college in a small coastal town in Maharashtra. This is public education doing something the private coaching industry could never do on its own. The coaching industry is very good at one thing. Teaching you to solve a known problem fast. That is what JEE and NEET reward, and I say that with love because I was part of that world. But an olympiad paper does not have a known type. There is no shortcut chapter. There is no formula sheet that saves you. You have to sit with a problem you have never seen and think. That is a completely different muscle. And a government funded centre has been quietly building it in Indian teenagers for 27 years. So yes, be proud. Loudly. HBCSE also shared that around 64 percent of India's olympiad medallists go on to do a PhD. But only about 32 percent of medallists end up settling in India. I do not say that to spoil the moment. These kids owe the country nothing. They earned every option they have. But it should tell us something. We are excellent at finding this talent. We are excellent at training it. We are still not great at giving it somewhere worth staying. Congratulations Kanishk, Riddhesh, Rishit, Shresth and Svarit. This is one of the best things an Indian did this year and most of the country will never hear about it.
🇮🇳 India Tops the World at the 56th International Physics Olympiad 2026! 🥇🥇🥇🥇🥇India's young physicists deliver an extraordinary performance at #IPhO2026 in Colombia. All five members of the Indian team win Gold Medals. 1/3 @PMOIndia @DrJitendraSingh @HBCSE_TIFR @TIFRScience
281
3,151
216
10,974
750,927
Bargava retweeted
This is an incredible blog post. Peak “centaur chess” phase of AI coding.
Jarred Sumner
@jarredsumnerJul 8
Rewriting Bun in Rust bun.com/blog/bun-in-rust
5
9
177
41,138
My skills repo has 160K stars, 7.5m downloads... ...and no tutorial. So, here it is. Watch me walk through the essential skills: - /grill-with-docs - /to-spec - /to-tickets - /implement - /code-review It's the whole flow, end-to-end. Enjoy:
121
641
55
7,371
594,591
Bargava retweeted
Replying to @Nithin0dha
Couple of nights ago, I had dinner at an event where I shared the table with someone from Zerodha. He's been working there for a while now. He had no air of working in India's No.1 brokerage firm. At some point, we discussed some of the competitors. He spoke well about them and things that they got right. No smirking, sarcasm, talking down others etc. This is so different to many VC-backed companies that I know - where the default state is toxic! @Nithin0dha you are an inspiration to all founders out there.
1
1
4
496
Bargava retweeted
Most people building agents don't realize they're just writing compilers That's why their harnesses are so brittle I'll show you the mental model that changed how I build, and debug, any agentic system from tinkering with 10+ harnesses I promise you will leave with a useful mental model which makes you 10x better at understanding any harness you ever see Doing a live breakdown July 7th 11:30 AM ET / 9 PM IST, link below
1
4
26
2,287
Bargava retweeted
Hall & Knight Series the meta theme is to to teach concepts every week and instill a culture of "not being scared of math" in kids / adults ... in some cases using AI tools to explain things more elegantly ( like Python code to understand graphs ) to treat Math + AI as a hobby rather than something you vanquish or compete on link in comments 1/2
3
3
34
2,764