Gaurav Sharma

I just open-sourced ContextDB - a unified context layer for AI agents. It replaces the patchwork of Pinecone + Redis + Postgres + glue code that every agent team builds with one system that actually understands memory. 82 tests passing. <5ms search latency. Apache 2.0. Thread on why this matters:

Brilliant demos 🤌🏼 What’s great & immediate win: Will open up gazillions of b2c opportunities- from especially around self-help apps, mental health/loneliness/virtual companionship related services for both young & old, gaming, personal ai agents will feel more personal. And, voice-controlled hardware (*cough* apple siri *cough*) What this is not for: This is not elevenlabs killer in b2b ai voice agents domain. Having helped close to 500+ businesses with globally with deploying voice ai agents in production - I’m pretty sure that businesses & their end customers will any day prefer a clear distinction b/w ai and human. The level of humaness Dessa brings on the table is kinda creepy in b2b’s perspective. The other challenge is - the brain & limbs aka access to tools ( and actions/webhooks) Voice AI Agents serving businesses have - knowledge bases - SOPs - complex if & else conditions - compliance requirements - integrations - triggers & post conversation actions - context management (not just during the call but from previous conversations & beyond) And, I have seen pure speech to speech model struggle with that. Even if we solve that the way Moshi solved it via MoshiRag, the next limitation comes via - limitation on the Talker leg of the model with limited number of voices and no/limited support for cloned voices - training truly S2S model (which hasn’t seen a single line of text) on multiple languages along with fluent cross-mixing becomes another massive pain point wrt deploying S2S for complex use cases in B2B Having said that - looking at the cracked team @dessa with @jack’s support - I’m sure they are going to absolutely kill it. Excited to get access to the APIs & play around 👍

As ai-powered threat actors & scripts are going ape shit finding and exploiting security loopholes across systems and frameworks, friendly reminder to minimize your loses (especially for new vibe coders): - put tight budgets and expiry dates on your temporary keys (be it LLM or any other api keys) - rotate your production keys, set budgets, set budget usage reminders for yourselves - don’t paste your keys as plain text in any of your code agents - double check your repos if they are not making any env variables public - double check gitignore & map files for npm packages - they should not be publicly accessible - if your vibe coded app is storing customer data, make sure you are not saving any PII & sensitive data in plain text in db. Ask your coding agent to confirm this Things are only going to get messy and crazy in coming days. So, let’s be prepared.

Ummmm, whatttt? @cursor_ai

Blessing in disguise for those who always wanted to start something but golden handcuffs didn't allow - come back and build from India, for the world. If that doesn't work out, so many amazing businesses are being built in all parts of the country, across multiple industries - Tech/AI, Manufacturing, Media, Ecom, Space & so on - join one and help them grow. AGI is anyways going to take everyone's job in the next 5-10 years 🫠

💔💔💔 lost my mentor, my big bro.

EU has probably done way more for Open AI than US government (when they blocked Claude Fable) ;) Reason I'm saying this is because ChatGPT has now officially been classified as a VLOSE, a Very Large Online Search Engine, by the European Commission. (refer EU Comissions post) The simple version: ChatGPT crossed 45M monthly users in the EU, which puts it into the stricter bucket of the Digital Services Act. OpenAI now has additional obligations around systemic risk, transparency, audits, illegal content, minors, mental wellbeing, fundamental rights, elections and public security. It has four months to comply with the additional VLOSE requirements. For businesses using ChatGPT, nothing suddenly breaks tomorrow or even in Jan 2027. It's more of a work to be done by OpenAI team. And, You don't need to ban it and someone using ChatGPT to rewrite an email is not suddenly a compliance event. But I do think this is a good moment for European companies to look at how quickly "we use ChatGPT" has turned into something much bigger. We started with rewriting emails, summarising documents and generating ideas. Now support teams are using AI to read tickets and draft customer responses. Sales teams are listening to every call, updating CRM records, researching prospects and writing follow-ups. HR teams are answering employee questions from internal policies. Ops teams are qualifying leads, booking appointments and generating reports. Legal is analysing contracts. Finance teams are experimenting with invoice checks, refunds and approvals. A simple way to think about it: - Using ChatGPT to write a better outbound email? Carry on. - Using it to summarise a meeting or brainstorm an ad campaign? Carry on. - A salesperson asking it to research a prospect before a call? Probably carry on. - A developer using it to explain some code? Again, probably not where I'd spend governance energy. But now change the scenario slightly: - Your support bot independently answers 30,000 customers a month using private customer data and internal documentation. I'd rethink how that is set up. - An AI agent reads Salesforce, decides which leads matter, sends the emails and changes CRM stages without a human looking at it. Worth rethinking. - An HR bot becomes the official source of truth for leave, compensation or employment policies across 5,000 employees. You probably want citations, versioning and a clear audit trail. - An insurance or lending workflow uses AI to recommend whether someone gets approved, rejected or escalated. Very different risk profile. - An agent can issue refunds, cancel accounts, change contracts or move money. You probably want explicit permissions and a human gate above a certain threshold. Your whole support operation is built around one model provider and switching away would take three months of engineering. That is also something I would rethink even if there were zero regulation involved. That's really the distinction. The more AI is just helping a person think, write or analyse, the less interesting this becomes. The more AI starts making decisions, talking directly to customers, touching sensitive data, changing systems or becoming impossible to replace, the more you probably want some control over what sits underneath it. And I don't think the answer for European businesses is to use less AI. Quite the opposite. Global competitors are not going to slow down because Brussels added another layer of regulation. If anything, the bigger risk is that European companies get so worried about compliance, data residency and vendor risk that they end up moving slower on AI while everyone else keeps accelerating. So the real question becomes: how do you keep using the best AI in the world without every important workflow becoming permanently welded to one provider? Maybe GPT is best for one part of a workflow, Claude for another and a cheaper open model is perfectly good for 70% of the boring work (usually it is). Maybe one customer’s data needs to stay in Germany. Maybe a regulated workflow needs full logs and human approval. Maybe the best model changes six months from now and you want to switch without rebuilding everything. None of this means your engineering team should now become an AI infrastructure company and start buying GPUs :) I think the better answer is going to be a Managed EU AI Stack: keep access to OpenAI, Anthropic, Google and open models, but have a managed layer that takes care of EU hosting where required, model routing, data boundaries, permissions, logging, evaluations, fallbacks, costs and switching models as the market changes. The goal isn't sovereignty for the sake of sovereignty. The goal is making sure European businesses don't have to choose between moving fast on AI and staying in control of it. This VLOSE announcement doesn't create that problem. It just makes it much easier to see. I'm spending a lot of time on this right now. If you're running a business in Europe and wondering which of your AI use cases are perfectly fine as-is and which ones deserve a rethink, happy to help or chat. DM or Comment. Bring 2-3 actual workflows. Happy to spend 30 minutes going through them together and literally bucket them into: keep as-is, add some controls or rethink the architecture.

Seen a lot of posts this week on the cost of Voice AI. Rajan Anandan from PeakXV said Voice AI could be the first $1B enterprise AI category in India outside coding, and that CIOs should not be paying more than ₹2/minute. Then a bunch of people broke down STT + LLM + TTS + telephony and landed somewhere around ₹1.5/minute all-in. The math is probably right. But having spent 10+ years in business telephony and working with businesses where the phone is literally their lifeline, I think we are simplifying the product a bit too much. Voice AI is not just STT + LLM + TTS. That is the underlying stack. It is not the product. Latency matters a lot, but not benchmark latency. What matters is how long the person on the other side actually waits before hearing a sensible response. I would want that somewhere around 500-700ms. Once every response starts taking a second or two, the economics may still look fantastic in a spreadsheet, but the conversation starts feeling weird. Same with conversation design and turn detection. Knowing whether someone has actually finished speaking or just paused to think/take a breath sounds like a small technical problem until you hear an agent constantly interrupt someone or wait awkwardly after every sentence :) A lot of Voice AI agents that sound great in demos still feel off because of this. Memory is another one. I don't think one sliding context window is really "agent memory". There are different types of memory. Things you want to remember from this call, things that should persist across calls, things about the customer/product/account, and also things the agent should eventually forget. I spent quite a bit of time thinking about this, and published a research paper earlier this year and open sourced some of our work here: papers.ssrn.com/sol3/papers.… github.com/atomsai/contextdb Coming back to cost, comparing Voice AI with human hourly cost is definitely one way of looking at it. In India you might be comparing against $1.5-2/hour, while companies outsourcing to the Philippines might be paying $5-10/hour. So yes, anything around or below that can produce ROI. But I wouldn't stop the calculation there. There is consistency, 24x7 availability, ability to suddenly handle 10x the calls, and the fact that when your pricing/product/policy changes you can retrain an AI workforce almost instantly rather than spending weeks retraining hundreds of people. More importantly, I think there are two very different ways to think about AI ROI. 1. How many people/costs can this replace? 2. How much more business can I do because this exists? I find the second one much more interesting. We recently helped one of our customers make around $6M in additional revenue largely by taking missed calls close to zero. They are also supporting more than a million calls a year without having to proportionally grow the team. Once that happens, the discussion around whether the agent costs ₹1.5/minute, ₹2 or ₹2.5 becomes much less important. In this case we actually charged a part of the upside. And this is probably the part I am most excited about. Think about a small clinic, car dealer, plumber, real estate broker, D2C brand or pretty much any small business. Earlier, spending more on marketing also meant: shit, who is going to handle all these extra calls? Who is going to answer at 9pm? Who will follow up with the leads? Who will book the appointments? What happens on Sunday? Do I hire two more people before I even know if the extra demand will come? So a lot of small businesses probably under-invest in growth simply because they don't have the people to service the demand. Voice AI changes that. ₹2/minute, ₹5/minute, even ₹10/minute... you suddenly have this polished, trainable, 24x7 gig worker who can pick up every call, learn a new product/service quickly and help you grow. Small businesses can think much bigger and be much more aggressive about growth without first having to build a big team. That, to me, is the real Voice AI story. Not that we managed to make a phone call 30 paise cheaper. But that millions of businesses that were constrained by people, hiring and operating hours can suddenly service a lot more demand without breaking the bank. There is another part of this conversation that I haven't seen discussed enough. AI agents can still go rogue. You can have the cheapest STT, LLM and TTS stack in the world, but what happens when the agent gives the wrong commitment to a customer, violates a compliance rule, applies the wrong discount or simply starts doing something stupid across thousands of calls? Evaluation, guardrails, tracing and qualitative review of these agents are going to become a pretty big part of the stack. Saving 20-30 paise/minute while ignoring this can turn out to be a very expensive optimisation. And this is also why I think the eventual revenue model for a lot of AI agents becomes more interesting than just $/minute. Yes, there will be very large Voice AI API companies making money by powering billions of minutes. But as the underlying cost keeps falling, I think we will increasingly price the actual work being done. $5-10 per appointment booked. $0.99 per support query resolved. $ per qualified lead. $ per payment collected. $9/month unlimited calls for an elderly loneliness solving app $9/month for teaching parents english and so on. Endless business opportunities as it will become easier to underwrite near-unlimited usage as costs go lower. At SaaS Labs, we have powered 1B+ conversations through our platforms over the years, so we have accumulated a fair amount of scar tissue around telephony, voice and audio engineering. We have a few launches coming up around some of these ideas over the next few weeks. Anyway, this is more of a brain dump than a prediction :) If you are exploring Voice AI agents for your business, happy to spend 30 minutes and share what we have learned so far.

"Karnataka to run voice ai pilots with Elevenlabs" Le Sarvam: "Aisa kya hai usme, jo mujhme nahi hai" 🫠

Mind blowing for realz! 🤯 Meta’s decoding language out of raw brain noise. Accuracy scales log-linear with data. Same lesson keeps repeating: it’s a data problem wearing a model costume. We’ve spent years on the same fight, different channel. The 8khz phone line. Narrowband, compressed, brutal. STT, TTS & S2S rebuilt from the metal up for telephony, sub-second latency on the worst audio there is, tuned millions of chunks of real calls. If you’re building voice agents and the phone line is killing your quality or your latency, I want you to test what we’re building. DM me. Early access & $200 in free credits.

Whole Indirapuram (Ghaziabad) has been dug up for last few months with open deep pits (death guaranteed if someone falls into it), debris and sand everywhere, a road gets closed randomly without proper marking & people are seen taking u-turns & longer time to reach home - usual no damnnn given for common people. Today may be you @myogiadityanath ji or someone from your govt is visiting and local authorities have very swiftly cleaned up and covered the area (which ideally they should have done in the first place had they cared about common taxpayers instead of just politicians and senior officials) I thought 10 times before writing this because obviously speaking up the truth is not liked by many but I could not take it this time. I’m not sure if I’m pissed or heartbroken - after paying >Rs 100 crores in direct taxes in last 10 years both personally and via my business - I still can’t get the basics and it hurts big time when the required things get done overnight when a politician is visiting and is just going to pass through the area in his car in seconds. It’s just painful to see all this. Or, may be I’m just too tired with all this.

TIL "caffeinate -dims" - the ultimate shipping 24/7 hack 🙀

To the fellow founders 👇❤️