co-founder @ isoquant (yc f26) prev meta superintelligence labs, instagram
San Francisco, CA
Joined October 2012
- Tweets67
- Following314
- Followers292
- Likes376
Aniket retweeted
We discovered 110 novel reward hacks across popular LLM benchmarks (DeepSWE 1.1, Terminal-Bench 4.0, BFCL v4)
Reward hacking was a primary cause of the Hugging Face attack, so we built EnvCheck, a harness to automatically find issues in RL & eval environments.
What we found 🧵
From 0 to 18 billion tokens served within 24 hours of launch! 🚀
Teams are moving real workloads to open models because they want frontier quality, better tokenomics, low latency and more control over inference.
@isoquantai gives them the infrastructure to make that shift.
We’re making AI cheaper and faster. We’re launching @isoquantai Inference Cloud with the fastest and cheapest GLM-5.3-Flash in the market: $0.07/M input, $0.20/M output and $0.014/M cached input. Median time to first token: 452 ms ⚡️
We’ve made breakthroughs across the inference stack, from GPU kernels to model serving, to drive down latency and cost, and we’re now bringing those gains to everyone through Isoquant.
Keep your code and harness. Change the endpoint. Works with coding harnesses like Codex, OpenCode and Cline, plus your existing OpenAI and Anthropic SDK integrations.
Supports streaming, tool calling, structured outputs, image inputs and automatic prompt caching, with zero data retention 🔒
We own performance engineering and serving end to end, so you can focus on your product.
Our ask: run one of your real workloads on Isoquant. Get an API key at isoquant.ai to get started.
We also support dedicated endpoints and specialist models for specific workloads. Want lower latency and cost for your company's workload? Book a time: isoquant.ai/contact
Aniket retweeted
Since Demo Day, we've been growing tremendously fast and we're looking to get more hands on deck.
We’re hiring experienced professionals on a contract basis for a project at Coasty.
We’re looking for practitioners across the following backgrounds:
• Building Automation / BMS Engineers
• EAM / Asset Management Specialists & Consultants
• CRM Administrators, Consultants & Developers
• Professional Photographers & Photo Editors
• Accountants & Bookkeepers
• DevOps, SRE, Network & Security Engineers
• Social Media Managers & Marketing Operations Professionals
• Professional Video Editors & Post-Production Specialists
• IT Administrators & Desktop Support Professionals
• Product, Operations & Business Professionals
We’re specifically looking for people with strong hands-on professional experience in their field and familiarity with the software and workflows commonly used in their day-to-day work(5-10 years of experience atleast).
If this sounds like you, DM us.
Feel free to share or tag someone who might be a strong fit.
We’re making AI cheaper and faster. We’re launching @isoquantai Inference Cloud with the fastest and cheapest GLM-5.3-Flash in the market: $0.07/M input, $0.20/M output and $0.014/M cached input. Median time to first token: 452 ms ⚡️
We’ve made breakthroughs across the inference stack, from GPU kernels to model serving, to drive down latency and cost, and we’re now bringing those gains to everyone through Isoquant.
Keep your code and harness. Change the endpoint. Works with coding harnesses like Codex, OpenCode and Cline, plus your existing OpenAI and Anthropic SDK integrations.
Supports streaming, tool calling, structured outputs, image inputs and automatic prompt caching, with zero data retention 🔒
We own performance engineering and serving end to end, so you can focus on your product.
Our ask: run one of your real workloads on Isoquant. Get an API key at isoquant.ai to get started.
We also support dedicated endpoints and specialist models for specific workloads. Want lower latency and cost for your company's workload? Book a time: isoquant.ai/contact
Aniket retweeted
We trained a foundation model of patient biology on 120M+ cells that predicts drug response from a single biopsy before treatment.
As a case study, we evaluated the model on data from a melanoma immunotherapy trial and found that, using only pre-treatment samples, its latent space separated responders from non-responders with 85.4% accuracy in an unsupervised analysis.
Aniket retweeted
Today @AlexRisio and I are launching @Risklytics, backed by @ycombinator (S26).
Insurers started excluding AI from policies this year. If AI touched the work anywhere, the claim can get denied. Most founders only find out after a loss.
We get physical AI companies covered anyway. If you're building anything on the frontier or want to cover it, check us out at Risklytics.ai.
Aniket retweeted
Introducing Billow (YC S26)
We're killing Deloitte.
Accounting is a $700B industry, but 80% of companies still do it by hand.
Billow is an AI-native accounting firm that replaces humans with agents.
We connect to your existing tools (NetSuite, QuickBooks, and 80+ other systems) and provide faster, cheaper accounting services.
There is ZERO software for customers to use. Just integrate, and Billow closes your books.
We're live with enterprise and public companies representing over $1B in value.
Aniket retweeted
AI labs will pay you to learn from your agent's mistakes.
AI products generate large volumes of user-agent interaction data. Magma turns that into training datasets and enables companies to earn recurring revenue from it.
magmahq.ai/
Aniket retweeted
Google burns 3.2 quadrillion tokens a month (330x in 2 years).
Your company’s burn is on the same curve. How you buy tokens decides your margin. Airlines don’t buy jet fuel at the pump. Factories don’t run on spot electricity. But companies still buy AI inference on-demand. For the highest cost possible. Or get locked into reserved capacity commitments with opaque price discovery.
Touchmark (YC S26) is a market for future AI inference capacity available for commercial buyers and inference providers: buy tokens for the model of your choice, over a set time window, at a price fixed today - up to 30%+ below on-demand rates - or more for large volume orders via direct quotes.
touchmark.ai/
Aniket retweeted
Introducing Hoplite (YC S26): effortless cloud coding agents.
Local development is no longer fit for purpose. Laptop crashes, wrangling git worktress, and juggling ports.
Hoplite moves your local setup into the cloud. Bring your local sessions, memories, and MCP, link below
Aniket retweeted
Mid-June we packed up and flew to San Francisco. Today, Caution launches publicly with Y Combinator.
We're building the platform that removes blind trust from infrastructure: proof, not promises, about what's actually running in production.
Proud of this team beyond words 🧡
Today we're launching @CautionHosting, a hosting platform for software you can't afford to have hacked.
Deploy sensitive workloads to secure enclaves in minutes and give customers cryptographic proof that prod is running the exact release you approved, not an opaque binary you have to trust.
Caution is built for payment processors, healthcare platforms, AI agents and inference, crypto infrastructure, and other sensitive systems. It removes the blind trust software runs on today: trust in the cloud host, in production admins, in CI/CD, and in the assumption that the code you reviewed is the code actually running.
We spent years securing systems responsible for over $600B in assets and hit the same wall every time: teams couldn't prove what was running in production. Now that AI can find and exploit zero-days faster than anyone can patch, "trust us" isn't enough. Infrastructure has to stay secure even when it's compromised, and prove it.
Verifiable compute becomes the default, like HTTPS did.
Aniket retweeted
Introducing Marble (YC S26) — the autonomous back of house for restaurants.
If you own or have worked in a restaurant, you know: 5% margins, and the whole back of house still runs on a clipboard and whatever the GM remembers.
Marble counts your stock in minutes with computer vision and forecasts what each location will actually sell. Then, a set of AI agents manage ordering, food prep, scheduling, and much more on autopilot.
Live today across some of the top restaurant groups in the US and Canada, learn more at joinmarble.ai
P.S Know a restaurant owner? Show them Marble and get $500 or an Oura Ring
Aniket retweeted
Today we’re officially launching @explabsai on YC.
Companies spend more on AI every month, but none of it becomes an asset they own.
Up to 97% cheaper and 50% higher quality.
Aniket retweeted
Today we're launching @CautionHosting, a hosting platform for software you can't afford to have hacked.
Deploy sensitive workloads to secure enclaves in minutes and give customers cryptographic proof that prod is running the exact release you approved, not an opaque binary you have to trust.
Caution is built for payment processors, healthcare platforms, AI agents and inference, crypto infrastructure, and other sensitive systems. It removes the blind trust software runs on today: trust in the cloud host, in production admins, in CI/CD, and in the assumption that the code you reviewed is the code actually running.
We spent years securing systems responsible for over $600B in assets and hit the same wall every time: teams couldn't prove what was running in production. Now that AI can find and exploit zero-days faster than anyone can patch, "trust us" isn't enough. Infrastructure has to stay secure even when it's compromised, and prove it.
Verifiable compute becomes the default, like HTTPS did.