Cherag Aroraa
San Jose, California, United States
2K followers
500+ connections
Activity
2K followers
-
Cherag Aroraa liked thisCherag Aroraa liked thisPrompt vs. Loop engineering, clearly explained! (why the creator of Claude Code stopped prompting it) These are Boris Cherny's exact words: "I don't prompt Claude anymore. I have loops that are running. My job is to write loops." That's the whole shift in one sentence, and most of us are still living in the first half of it. 𝗣𝗿𝗼𝗺𝗽𝘁 𝗲𝗻𝗴𝗶𝗻𝗲𝗲𝗿𝗶𝗻𝗴 𝗶𝘀 𝘁𝗵𝗲 𝗺𝗲𝘀𝘀𝗮𝗴𝗲. The model has no memory before the call, so the prompt carries everything it needs: role, instructions, examples, format. It reads the message, generates a response, and stops. Nothing repeats unless you send another prompt. So the skill is composing that one input well, and knowing which ingredient is letting the output down when it falls short. 𝗟𝗼𝗼𝗽 𝗲𝗻𝗴𝗶𝗻𝗲𝗲𝗿𝗶𝗻𝗴 𝗶𝘀 𝘁𝗵𝗲 𝘀𝘆𝘀𝘁𝗲𝗺. A loop doesn't end when the model stops talking. It ends when the job is actually done, and those two things get confused constantly. A coding agent writes some code, glances around, sees progress, and announces it's finished. The tests still fail underneath it. Every loop runs the same six lines: call the model, run the tool calls it asks for, feed the results back into context, repeat until there's nothing left to call. That part is already solved. Nobody is competing on the while statement. What's hard is everything wrapped around it. Knowing when to stop comes first. The model saying "I'm done" isn't proof, so what proves it is an automated completion check like passing tests, backed by hard caps on time and iterations. Keeping context clean matters just as much. Old tool outputs and dead ends pile up until the model's decisions get worse every turn, so context gets treated as a budget and summarized along the way instead of carried in full. Tools need the same discipline. A focused, non-overlapping set beats a hundred options, and every write needs to be safe to call twice, because loops retry. And something in there needs to be able to say no. Left alone, an agent agrees with its own work, so the fix is a test, a type check, or a separate model grading the output instead of the agent grading itself. Here's the part that ties it together. Prompting is you steering the agent move by move. Loop engineering is you building the system that steers it, then stepping back. _____ Share this with your network if you found this insightful ♻️ Follow me (Akshay Pachaar) for more insights and tutorials on AI and Machine Learning!
-
Cherag Aroraa reacted on thisCherag Aroraa reacted on thisThe San Jose Semaphore has been solved! The puzzle, created by Ben Rubin, features rotating discs at the top of Adobe’s Almaden Tower that hide a message through data points of bytes and numbered colors. The difficulty level? Perfect. At least, that’s what Brian Vincent said after solving the puzzle. How did he solve it? Watch his full journey and hear Ben Rubin’s reaction at https://adobe.ly/semaphore
-
Cherag Aroraa liked thisCherag Aroraa liked thisIt's been over a month since we released Muse Spark on the Meta AI glasses (Rayban Meta, Oakley Meta) and I've been seeing posts of people experiencing a shift in the AI performance, both in terms of capability but also conversationality. We saw these in our benchmarks and testing but it's sweeter when the end user experiences it firsthand. https://lnkd.in/eVCaKBm3 https://lnkd.in/e8gwmFfR Look forward to seeing what people start using the glasses for as the experiences become more complex and richer. It's been an honor to work on the posttraining for Muse Spark!
-
Cherag Aroraa liked thisCherag Aroraa liked thisA good week for open-weight LLMs. A bunch of nice models that run locally (even Gemma 4 12B only needs less than 16 GB of RAM). Ok, Nemotron 3 Ultra is a bit too large to run locally, but it comes with a seriously impressive performance: efficiency ratio, and will be an open-weight model hard to beat in this aspect for a long time. At a high level, Nemotron 3 Ultra is the large sibling of Nemotron 3 Super. The overall design is still a hybrid Mamba-Transformer MoE stack, but scaled to 550B total parameters with 55B active per token. The part I find most interesting is the Latent MoE idea (introduced in Nemotron 3 Super). In a regular MoE layer, the routed experts operate directly at the model width. In Latent MoE, the routed path is first projected down into a smaller latent space, the experts operate there, and the result is projected back up. For Super, this was 4096 -> 1024 -> 4096. For Ultra, it is 8192 -> 2048 -> 8192. So the 4x compression ratio stays the same, but the model is scaled up substantially. This is a nice example of architecture scaling by combining several efficiency mechanisms: Mamba-2, GQA, Latent MoE. As always, more details and higher-res figures in my LLM Architecture Gallery.
-
Cherag Aroraa reacted on thisCherag Aroraa reacted on thisHi Thrilled to share that our research paper, “Efficient Context-Aware Corpus-Grounded Query and Prompt Suggestion Constrained by Creative Knowledge Graph”, has been accepted at ACM synthAI@SIGMOD 2026. This work explores how creative knowledge graphs and corpus-grounded retrieval can be leveraged to generate more relevant, context-aware query and prompt suggestions at enterprise scale. As AI assistants become increasingly integrated into creative workflows, grounding and contextual relevance remain critical challenges, and this research is a step toward addressing them. I’m excited to be attending SIGMOD 2026 in Bengaluru (Whitefield) from May 30 – June 6, where I’ll be presenting our work and engaging with the global data management and AI research community. A sincere thank you to my co-authors (Twisha Naik Vipul Jain Cherag Aroraa, Sudhakar Pandey Vedanth Subramaniam Eshan Trivedi ) collaborators, reviewers, and the broader Adobe engineering teams whose insights and support made this possible. Looking forward to the discussions, learning opportunities, and new connections at SIGMOD! #SIGMOD #SIGMOD2026 #Research #Databases #GenerativeAI #KnowledgeGraphs #LLM #LoRA #PromptEngineering #EnterpriseAI #AdobeResearch #DataManagement
-
Cherag Aroraa liked thisCherag Aroraa liked thisAnthropic just released Opus 4.8. (and the best feature isn't the model) The benchmarks are impressive. The honesty improvements are real. Fast mode is 3x cheaper. But what shipped alongside the model might matter more for how we actually build: Dynamic Workflows in Claude Code. Here's what they are, how they differ from subagents and agent teams (which already existed), and why they change the game for large-scale agentic coding. 𝗪𝗵𝗮𝘁 𝗮𝗿𝗲 𝗗𝘆𝗻𝗮𝗺𝗶𝗰 𝗪𝗼𝗿𝗸𝗳𝗹𝗼𝘄𝘀? Claude Code already had two multi-agent primitives before this release. Subagents are lightweight workers spawned from a main session. They do a focused task and report back. But they can't talk to each other, and the main agent still acts as the bottleneck for orchestration. Every result routes through one context window. Agent Teams (shipped with Opus 4.6) removed that constraint. Multiple Claude instances coordinate through a shared task list and message each other directly. But they top out at 3-5 teammates practically, sessions don't survive interruptions (if Claude crashes mid-task, the team is gone), and you still need to design the orchestration upfront. Dynamic Workflows sit above both. Instead of Claude holding the plan in its context window, it writes a JavaScript orchestration script. That script becomes the plan. A JS runtime executes it, fanning work across tens to hundreds of parallel subagents automatically. You describe the task. Claude decides how to split it, how many agents to spawn, how to verify results, and what to report back. The orchestration logic moves from the LLM's memory into executable code. Claude's context window only ever sees the final converged answer. Not the intermediate results of hundreds of steps. 𝗪𝗵𝘆 𝘁𝗵𝗶𝘀 𝗺𝗮𝘁𝘁𝗲𝗿𝘀 → 𝗦𝗰𝗮𝗹𝗲. Up to 16 concurrent agents, 1,000 total per workflow. → 𝗔𝗱𝘃𝗲𝗿𝘀𝗮𝗿𝗶𝗮𝗹 𝘃𝗲𝗿𝗶𝗳𝗶𝗰𝗮𝘁𝗶𝗼𝗻. Agents tackle a problem from independent angles, other agents try to refute their findings, and the system iterates until answers converge. Subagents just report back. Agent Teams collaborate but don't adversarially verify. → 𝗥𝗲𝘀𝘂𝗺𝗮𝗯𝗶𝗹𝗶𝘁𝘆. Progress saves continuously. Interrupted jobs pick up where they left off. Agent Teams die with the session. → 𝗭𝗲𝗿𝗼 𝗼𝗿𝗰𝗵𝗲𝘀𝘁𝗿𝗮𝘁𝗶𝗼𝗻 𝗯𝘂𝗿𝗱𝗲𝗻. You describe the goal. Claude decides how to split work, how many agents to spawn, and how to verify results. 𝗕𝗲𝘀𝘁 𝗽𝗿𝗮𝗰𝘁𝗶𝗰𝗲𝘀 → Start with a scoped task to calibrate token usage before going full-scale. Workflows consume significantly more tokens than a typical session. → Enable auto mode so Claude decides when a workflow is appropriate vs. when a simpler approach works. → Use the `ultracode` setting (effort menu) to let Claude auto-trigger workflows. This also sets reasoning to `xhigh`. ____ Share this with your network if you found this insightful ♻️ Follow me (Akshay Pachaar) for more insights and tutorials on AI and Machine Learning!
-
Cherag Aroraa liked thisCherag Aroraa liked thisClaude Code does more than write code. How to build a full agent system in 6 steps: Written guide: https://lnkd.in/eG2-Fj6S Video guide: https://lnkd.in/eZrSQa2t Step 1: Install Claude Code ✦ Terminal, desktop app, or VS Code extension. ✦ Pick whichever surface fits how you work. ✦ Everything below it compounds. Full setup guide: https://lnkd.in/eMEky7te Step 2: Build context ✦ ~/CLAUDE.md is the onboarding doc ✦ Claude reads before every session. ✦ Every chat loads it automatically. Step 3: Build memory ✦ Every correction becomes its own .md file. ✦ The same mistake never lands twice. Step 4: Build skills ✦ One command fires a whole workflow across tools. ✦ No more retyping a 200-word prompt every time. ✦ /your-skill → Notion → Gmail → Drive → Output Step 5: Build agents Each agent gets one file and one job. ✦ Strategist (Opus) → analyses the task ✦ Builder (Sonnet) → executes the output ✦ QA Gate → fails anything below 95/100 Step 6: Run on autopilot Claude Routines run your team on Anthropic's cloud. Set the schedule. Walk away. ✦ Daily 8am UK → runs in cloud ✦ Weekly Monday → drops to Notion This isn't "using Claude." It's a system that works without you in the room. Repost ♻️ to help someone in your network. P.S. Which step are you currently on?
-
Cherag Aroraa liked thisGame onCherag Aroraa liked thisGenerate Design is now available in the Adobe Express AI Assistant! A world-class Layered Design Generation capability, powered by the Firefly Design model, turns a detailed prompt into a fully editable Adobe Express document in seconds. Every shape, text block, and image arrives as its own editable layer, ready to refine and make your own. This innovation was driven by our team in Romania through a cross-team collaboration spanning multiple geos. Proud to see this kind of partnership deliver real value for the millions of solopreneurs and small businesses who want to tell their story. #AdobeExpress #AdobeFirefly #oneteam #createthefuture #raisethebar
-
Cherag Aroraa reacted on thisCherag Aroraa reacted on thisAfter 𝟴 𝘆𝗲𝗮𝗿𝘀 𝗮𝗻𝗱 𝟯 𝗺𝗼𝗻𝘁𝗵𝘀, I’m closing one of the most meaningful chapters of my career. Leaving 𝗔𝗱𝗼𝗯𝗲 𝗮𝘀 𝗣𝗿𝗶𝗻𝗰𝗶𝗽𝗮𝗹 𝗦𝗰𝗶𝗲𝗻𝘁𝗶𝘀𝘁 in the Search & Discovery org wasn’t easy. I had the privilege of leading the 𝗜𝗻𝘁𝗲𝗻𝘁 𝗔𝗜 team and shipping features that touched millions across Adobe 𝗣𝗵𝗼𝘁𝗼𝘀𝗵𝗼𝗽, 𝗘𝘅𝗽𝗿𝗲𝘀𝘀, 𝗮𝗻𝗱 𝗟𝗶𝗴𝗵𝘁𝗿𝗼𝗼𝗺. Deep gratitude to everyone I collaborated with. 🙏 At Adobe, I learned connecting 𝘂𝘀𝗲𝗿 𝗶𝗻𝘁𝗲𝗻𝘁 to 𝗱𝗼𝗺𝗮𝗶𝗻 𝘀𝗽𝗲𝗰𝗶𝗳𝗶𝗰 𝗸𝗻𝗼𝘄𝗹𝗲𝗱𝗴𝗲 leads to delightful user experiences. We applied that to creativity. Now, I’m applying it to something even more fundamental: 𝗛𝗲𝗮𝗹𝘁𝗵 𝗮𝗻𝗱 𝗪𝗲𝗹𝗹𝗻𝗲𝘀𝘀 About 18 months ago, I started a personal experiment. I was a busy professional, borderline overweight, and like many of us — aware something needed to change, but unsure where to start. By analyzing my own data with tools and health agent (shoutout to Cursor/Windsurf), I did something I’d never managed before: I truly understood where 𝗺𝘆 𝗵𝗲𝗮𝗹𝘁𝗵 𝗶𝘀 𝗵𝗲𝗮𝗱𝗶𝗻𝗴, and which parts needed attention! I designed a plan around a 𝟱-𝗵𝗼𝘂𝗿 𝗲𝗮𝘁𝗶𝗻𝗴 𝘄𝗶𝗻𝗱𝗼𝘄, 𝗹𝗲𝗮𝗻 𝗽𝗿𝗼𝘁𝗲𝗶𝗻𝘀, 𝗮𝗻𝗱 𝗮 𝗵𝗼𝗺𝗲-𝗯𝗮𝘀𝗲𝗱 𝗺𝗶𝘅 𝗼𝗳 𝘀𝘁𝗿𝗲𝗻𝗴𝘁𝗵, 𝗰𝗮𝗿𝗱𝗶𝗼, 𝗮𝗻𝗱 𝗳𝗹𝗲𝘅𝗶𝗯𝗶𝗹𝗶𝘁𝘆. I added mindfulness and meditation, which made me aware of how 𝗻𝘂𝘁𝗿𝗶𝘁𝗶𝗼𝗻, 𝗺𝗼𝘃𝗲𝗺𝗲𝗻𝘁, 𝘀𝗹𝗲𝗲𝗽, 𝗮𝗻𝗱 𝗲𝗺𝗼𝘁𝗶𝗼𝗻𝘀 𝗮𝗿𝗲 𝗱𝗲𝗲𝗽𝗹𝘆 𝗶𝗻𝘁𝗲𝗿𝗰𝗼𝗻𝗻𝗲𝗰𝘁𝗲𝗱. The results showed up everywhere: • Clinical markers 𝗶𝗺𝗽𝗿𝗼𝘃𝗲𝗱 • Mental clarity 𝘀𝗵𝗮𝗿𝗽𝗲𝗻𝗲𝗱 • I can now 𝘀𝗸𝗶 𝘄𝗶𝘁𝗵 𝗺𝘆 𝗱𝗮𝘂𝗴𝗵𝘁𝗲𝗿 𝗳𝗼𝗿 𝗵𝗼𝘂𝗿𝘀 without getting tired Health agent didn’t give me a generic plan. It helped me understand me. That belief led me to my next chapter. I’m excited to share that I’m joining 𝗣𝗿𝗲𝗳𝗳𝗲𝗰𝘁 𝗮𝘀 𝗖𝗧𝗢. Preffect is building 𝗽𝗲𝗿𝘀𝗼𝗻𝗮𝗹𝗶𝘇𝗲𝗱 𝗔𝗜 𝗲𝘅𝗽𝗲𝗿𝗶𝗲𝗻𝗰𝗲𝘀 across nutrition, physical activity, sleep, and emotional wellbeing — treating health as the 𝗶𝗻𝘁𝗲𝗿𝗰𝗼𝗻𝗻𝗲𝗰𝘁𝗲𝗱 𝘀𝘆𝘀𝘁𝗲𝗺 it truly is. The gap between knowing 𝘄𝗵𝗮𝘁’𝘀 𝗴𝗼𝗼𝗱 𝗳𝗼𝗿 𝘆𝗼𝘂 𝗮𝗻𝗱 𝗱𝗼𝗶𝗻𝗴 𝗶𝘁 can finally be closed — through guidance grounded in your own data. If you’re curious, I’d love for you to try the app and share your feedback. We’re building this thoughtfully and would value input from this community. If you’re working at the intersection of AI and health — or thinking about it — let’s connect. I’m excited to engage more deeply with the SF startup ecosystem and collaborate with like-minded builders. #preffect #health #wellness #ai
Experience
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
Education
Licenses & Certifications
Projects
-
Traffic Object Detection
-
• Created a pipeline to train YOLO model on custom traffic scenarios from LG simulator.
• Implemented ROS based application which classifies objects such as Car, Pedestrian with bounding boxes.
• Refactored the LG simulator and ROS Bridge and integrated this new Model to showcase bounding boxes on simulator scenarios.
Other creators -
-
Admission rate prediction
-
See project• Created a patient readmission predictor using Pickle & Kafka to ship the models to the runtime server.
• Used algorithms such as XGBoost, KNN and Random Forests to create models and evaluate performance.
• Performed data cleansing and data preprocessing on dataset containing medical records.
-
OrganChain (Organ donation)
-
See project• Developed a resilient blockchain application which maintains an organ sharing ledger in NodeJS.
• Containerized the block chain network over AWS. Used Hyperledger Fabric to store the organ lifecycle into a block.
• Implemented DAL to store data in MongoDB and reactJS for front end application.
• Analyzed hospital and donor information to provide closest organ donation location.
-
Easy-Gift (Ruby on Rails, MongoDB, Bootstrap, Jquery, Facebook API)
-
Built application to provide gift options to the users through crowdsourcing. Application allowed users to post the gift request so other users can comment and provide a gift option. Business users within the system subscribed to the specific tags will get email notifications so they can send private message to the users directly.
• Implemented an application which allows users’ to post a gift suggestion request for their loved ones (Similar to stackoverflow.com).
• Implemented chat…Built application to provide gift options to the users through crowdsourcing. Application allowed users to post the gift request so other users can comment and provide a gift option. Business users within the system subscribed to the specific tags will get email notifications so they can send private message to the users directly.
• Implemented an application which allows users’ to post a gift suggestion request for their loved ones (Similar to stackoverflow.com).
• Implemented chat, private messaging and private post features.
• Used MongoDB to store the data.
• Integrated Facebook API so users can share their gift ideas with their friends.
• Connected local business with the users.Other creators -
-
Splunk Data Analysis - Splunk, Yelp Data set
• Configured and deployed Splunk cluster of indexer, forwarder, and master.
• Analyzed different metrics on Yelp Data like users with most reviews (Average query time : 30 sec)
• Created visualization using Splunk and IBM many eyes.Other creators -
-
Uptacks.net - ASP .NET, C#, MS-SQL, Facebook
-
• Developed a web application allowing users to pin their favorite bookmarks with images and videos.
• Implemented backend in C# OOP and front end in Javascript, Jquery, CSS Twitter Bootstrap, HTML5.
• Integrated Facebook API for users for quick access.
• Implemented social networking features such as following, follower, news feed, comments and making pins favorites. -
Twitter Data Analysis - HiveQL, Hadoop, HDFS, CentOS 6.2
-
• Configured and deployed system with Hadoop cluster of 5 nodes (1 master & 4 slaves). (Average query time: 1 minute).
• Data from Twitter was fed in HDFS and migrated to Hive. (Average query time: 1 minute).
• Analyzed interesting co-relations and trends in data using HiveQL for Hadoop.
• Optimized the tables to handle 15 GB data from Twitter spanning over 24 hours.Other creators -
-
HealthyBytes Web Application Testing -Selenium, Unit Testing, Automation UI Testing, Performance Testing
-
• Create test plans for functional, scalability and performance testing of cloud based web application
• Develop test case based on the test plans and execute it to test the cloud based software.
• Develop and execute testing automation scripts for all the test cases.Other creators -
-
Testing as a service – Selenium, Unit Testing, Automation, Performance Testing
-
• The goal of the project was to enhance the given TaaS tool, known as HBSA, and perform function and performance testing for a selected Software-as-a-Service (SaaS).
• Generated Selenium scripts for automation and used JMeter for performance testing.
-
Mentor Web Application - MVC Framework, PHP, MySQL, Skype API
-
• Developed a social networking application for the interaction between mentors and mentee using Codeigniter.
• Developed a platform for SJSU students so they can get tutoring online over chat and video calls.
• Implemented and integrated chat feature and Skype API for video calling. Created test cases & requirement documents. -
Automaker Database Application – Oracle 11g, Java, SQL
-
• Developed an Automaker Database application for car vendors to display the cars and its parts from different automakers.
• Optimized queries to reduce the query time. Used Java and Oracle 11g database to store the data.
• It requires knowledge of Oracle optimizer metrics & procedures.
Other creators -
-
iSpartan Social Networking Android Application & Website – SQL Database, PHP, Android SDK
-
• Developed an android application using Android SDK following Mobile design principles and methods.
• Developed social networking website using PHP, MySQL, and MVC framework.
• Added features to locate hot spots near campus, post on friend’s wall and search friend in friend list.
• Integrated Google Maps API with the application and website to locate friends’ location. -
Airlines Reservation System – Java, J2EE, Netbeans IDE, MySQL, JUnit
-
• Developed 3 tier application for Airline Reservation System.
• Developed MySQL based database, created service to provide functionality.
• Created JSP/Servlets based GUI client. Performed JUnit testing on server side methods/functions.
• Tested the code for performance by applying Caching, Connection Pooling concepts and performing tests using Jmeter.
• Performed scalability test for concurrent threads using Jmeter.
Honors & Awards
-
Dean's List
-
(For 3.5+ GPA over 4 in two consecutive semesters)
Recommendations received
10 people have recommended Cherag
Join now to viewExplore top content on LinkedIn
Find curated posts and insights for relevant topics all in one place.
View top content