Krish Dasgupta

It was fun presenting at NODES 2025 ! Thanks @neo4j @JulesHitz

@anthonyronning @lateinteraction @TryMapleAI @axonvertex_ai Shall DM👍🏼

@officialKrishD @lateinteraction @TryMapleAI @axonvertex_ai Yes let’s definitely chat more. Lots of shared interest here.

okay i like half understand RLMs now and it's sick

@anthonyronning @lateinteraction Anthony - want to present something together on confidential compute ? Can DM you and we take it from there. There’s a lot. Seems like @TryMapleAI has a lot of points of interests that @axonvertex_ai can contribute towards.

@officialKrishD @lateinteraction I would love that!

@anthonyronning @lateinteraction Yes, It is absolutely much needed 😊🤞🏼. Also, am presenting part of this work in AppliedAISummit. I will share with you once the talk and demo are published.

@officialKrishD @lateinteraction This is sick! The graph component seems very powerful here. Thank you for sharing!

On the conversational part, Anthony, you can try these reasoning approaches (I am already using a few of them in informed clinical decision making with human in the loop intervention, where the human serves as RLHF feedback): 1. Create context from the conversation and map the key elements as graph nodes. 2. Generate vector embeddings for these nodes and build a hybrid vector with graph structure. 3. From the graph embeddings, retrieve and traverse relevant paths to produce true contextual reasoning for the specific query or task. General reasoning is very different from contextual reasoning. For example, in my medical use case: if a medical profiler is asked, “Is smoking harmful for the person?”, it will simply say yes based on general knowledge. Contextual reasoning, however, looks into the patient’s full path, for instance, if the patient is a chain smoker who has recently been diagnosed with COPD, it determines that it is now extremely dangerous and the risk is very high. Contextual reasoning is the way upwards on contextual graph. Also DSPy will play a vital role.

@officialKrishD @lateinteraction Yeah I'm already thinking about how to use it with our DSRs-based persistent agent whenever it needs to reason across more complex conversational history. Like reasoning with human behavior across time.

@lateinteraction @anthonyronning Anthony - wait till you think to use RLM in reward policies and see the outcome.

@anthonyronning haha what happened

AI startups - for every million dollars raised, plant 1/10 of those many million saplings and trees.

@jayitabhattac11 @claudeai Me ( again ) : Claude, use my github repo, my cv, my last 1 year tweets and tech talks, and synchronize the talk according to my context and knowledge. No more, no less. And yes, DO NOT add taste and flamboyance by yourself.

Claude: vibe deployed app Tech Lead: explain how it works Me: @claudeai summarize what you did ;)

What specific areas are you planning to invest in the next cohort - especially in healthcare ? Which focus areas do you see the traction to grow and will have large market pie in healthcare ? @ycombinator , can you help ?

Had a short call with a senior VP of a well known Indian IT company regarding Agents and Agents Memory. Take : The seniority was only in age. Turns out, the VP responsibility was only to blabber random “AI terms” by memorising terms from a bot. Last coded when asked just like that - 12 years back. Turns out - VP is responsible to drive the agent development ecosystem. No wonder we have so minimal AI understanding with colossal gaps. Young Folks graduating recently - please don’t choose a person as your mentor / company / manager who themselves have never done anything but can blabber. It will destroy you in long term. Analogy - Certain VCs who have never ever broken a sweat or coded anything in life are directing your startups and choices. Duds.

@jpatel41 How do I reach out to you @jpatel41 ? Your DMs are closed. Want to share the earlier draft of Agentic guardrails that I have been working on.

The ability to project out just 3-6 months longer than other companies could become a massive superpower for any company. The crazy part is that this doesn’t even require you to be a visionary. It just requires recognizing that there is massive information asymmetry and actively attempt to tap into that information asymmetry. The uneven distribution of the trends that have already manifested but not yet obvious to everyone could become a material strategic execution advantage for any company. But it requires constantly staying current on all trends regarding AI which can be exhausting, but more necessary now than ever before. Last thing I’ll say is that it is so tempting to draw historical analogies of the current environment. But the thing that is counterintuitive is that with AI there are literally “no priors”.

@TomasMann1878 Wonderful !

Tools I actually use as a research engineer + SaaS builder at night: 𝗥𝗲𝘀𝗲𝗮𝗿𝗰𝗵 - You (the ideas still come from the human, sorry) - Perplexity - NotebookLM - Deep Research 𝗔𝗴𝗲𝗻𝘁 𝗗𝗲𝘃 - OpenClaw - Pure Python mostly - ClawBench (run your agents against each other - surprisingly revealing) - Weights & Biases 𝗠𝗲𝗺𝗼𝗿𝘆 - Mem0 - Zep - Chroma - Pinecone 𝗖𝗼𝗱𝗶𝗻𝗴 - Claude - Codex - Antigravity 𝗜𝗻𝗳𝗿𝗮 / 𝗢𝗿𝗰𝗵𝗲𝘀𝘁𝗿𝗮𝘁𝗶𝗼𝗻 - Vercel - Runpod - Convex/Supabase 𝗪𝗿𝗶𝘁𝗶𝗻𝗴 / 𝗗𝗼𝗰𝘀 - Notion AI - Obsidian 𝗠𝗼𝗻𝗶𝘁𝗼𝗿𝗶𝗻𝗴 - Langfuse - Helicone - PostHog - Promptfoo The job is still 80% reading papers and 20% crying about context windows. Save this. Build something.

For anyone who thinks quick commerce delivery is about innovation, it is about labour economics. As simple as it gets.

NY and Seattle passed a local law for introducing a bill of minimum pay for app based delivery workers ( 17-20$) US market knows better that to account for faster delivery they cannot employ but actually create partners ( gig workers ) to ascertain the delivery in quick commerce. Companies if bound to pay by local law- will juice out the startup / quick commerce company- India - many startups run in losses to still accommodate the win in ( market reachability ) quick commerce. So, it is economics- plain old labour economics and nothing of innovation. Refer : nber.org/digest/202603/…

@AditiS90 @ShwetaRKohli Well, it is economics -rather the gig economy to say that allows the labour cost equated in this process in India.

The US is discovering 1-hour delivery...India has already leapfrogged to 10 minutes. :)

@GoDaddyHelp Ok this helps. But, if there are outages - at least notify - I almost missed a deadline !

@officialKrishD Glad it's back up and running. We understand the frustration, and appreciate you letting us know. If anything comes up again, our Support team is available 24/7 at godaddy.com/help/contact-us. ^CG

Damn You @GoDaddy ! If I miss a deadline , you better be ready ! Damn You ! What the actual hell ????? @GoDaddy

@GoDaddy No emails. No messages. just auto corrected after 1 hour. Come on ! Be better. It has started working now.

@GoDaddy @GoDaddy What the Hell ? Is there no 24/7 support ?????

@askalphaxiv @gowthami_s Beautiful creation !!!!

Introducing MCP for arXiv Let your research agents stand on the shoulders of giants Fast multi-turn retrieval, keyword search, and embedding search tools across millions of arXiv papers 🚀

Me and HDFC - a day in a life of customer call. ( called a number and the post the option of English as the main language ) Bot : Welcome to HDFC bank. In few words can you tell me what is your issue? Me : I have some pending remittances which are not cleared. Bot : Sorry , I did not get you. In few words can you tell me what is your issue? Me : I… Bot : Sorry , I did not get you. In few words can you tell me what is your issue? Me : Remittance. Pending . Issue. Help. Bot : I see you have queries for pending remittances. To help you better, I will transfer this call to an agent - Basically the bots are completely intent driven. Choose the words - they match and process. VC money will be used to solve this issue . Lol