Activity
957 followers
-
Shailesh Ahuja shared thisI have been holed up for the last six weeks, rebuilding Jenni AI chat from the ground up. It now wins in 76% of the benchmarked runs compared to the old agent. If you are working to replace frontier models with smaller, cheaper ones while maintaining or improving quality, or are interested in practical applications of loop engineering, this will be a great read.
-
Shailesh Ahuja shared thisThis is my favourite LLM chart. Not every AI feature needs the best model. The Pareto Frontier defines which models are most intelligent at their price point. While most model releases focus on beating the top benchmarks, very few advertise whether they are best for their price. Finding the Pareto Optimal model for a specific feature is a lot of work. It requires evals, rigorous prompt iteration, and LLM Ops. But that's the difference between spending millions or thousands of dollars on AI tokens.
-
Shailesh Ahuja shared thisI have been thinking about what AGI will mean for society and the future. After weeks of research and reflection, I was able to gather my thoughts on this futuristic technology that's almost here. My thesis is that the biggest challenge we face in the Post-AGI world is not the technology disruption itself, but how we handle the transformation and distribute AGI's output in society. Read more here:
-
Shailesh Ahuja posted thisOne Big Bet. 20% Higher Conversion. In start-ups, especially in lean startups, we are always looking for those small changes that have an outsized impact on the customers. Those low-effort needle movers. However, there are times when we have to take big, bold bets, which are high effort, high risk, but potentially high reward. Those are the tasks that give the most fulfillment when they finally execute well. This task was one of those. We bet that if we index all open-access papers in-house, and let our LLMs loose on them with RAG, it will improve our users experience with far richer context. Having full-text PDFs helps users not only get suggestions from all parts of the paper, we also let them trace exactly where it's cited from by adding a link to the exact quote in the full text. It took us many months of effort to figure out the best way to do this at scale. We had so many open questions. How do we collect these papers? What's the best embedding model to use? What's the best way to generate chunks? How do we ensure the vector search is performant at this scale? After all these questions were resolved, the execution was equally long. We spent multiple weeks just waiting for our scripts to finish and tens of thousands of dollars to generate embeddings for these papers. We had some good signs it will help from smaller scale experiments, but we didn't know for sure until we flipped the switch on. And the results were great: 1. 20% increase in conversion rate 2. More relevant citations - Auto-complete and citation search see 10% higher acceptance rate 3. More traceable citations - Directly linking to PDFs improved PDF viewer feature adoption by 3x 4. Intrastructure to build more useful agents Data processing that takes multiple weeks cannot be trusted with AI. All painstakingly human-managed work in the hope of something better, and so glad it worked out!
-
Shailesh Ahuja shared this🚀 Scaling jenni.ai to One Billion Vectors I just published a deep dive on how we scaled Jenni's vector database architecture from 100M+ to nearly 1B managed vectors. It powers autocomplete, agent workflows, and citation search across millions of research documents. This post covers: ▶️ Options we evaluated: Qdrant, Pinecone, Vertex AI, Milvus/Zilliz, pgvector & Turbopuffer ▶️ Why Qdrant ended up being the most pragmatic choice ▶️ How we optimized costs using quantization and oversampling + rescoring ▶️ Challenges faced scaling Qdrant If you're working on RAG, vector databases, or scaling LLM systems, hope this helps! 👇
-
Shailesh Ahuja shared thisJust got this in the mail today. While OpenAI is a fraction of our usage, this "token" is a great reminder of how grateful I am to build an AI native product from this part of the world, a product that wouldn't have existed without the LLM revolution. Web development was mostly CRUD (boring) before, but now there are so many interesting UX challenges to help people drive most value out of LLMs. Creativity is much more valuable than skills or memory, more time is available to brainstorm and less time is spent worrying about repetitive execution.
-
Shailesh Ahuja shared thisAbsolutely love to see customer feedback like this. Our fully remote team has made huge improvements at a rapid pace, with ~4 engineers over the past year, and meeting 1-2 times per *week*. After some iteration, these are the expectations we landed on: 1. Send at least one Slack message daily. Most engineers actively collaborate on Slack, participating in 3-4 threads per day. These include code reviews, UX feedback, idea sharing, technical discussions, and more. 2. Maintain a solid 4 hours overlap with GMT+8 for real-time communication. Having quick back-and-forth exchanges is critical for async discussions. 3. Ship something daily - Whether it's a PR, notion doc or other deliverable. With the most advanced AI tools at our disposal, make meaningful progress every day. At the same time, we encourage engineers to solve problems from customer feedback proactively and share prototypes to get feedback ("Show And Tell"). Our product team ensures engineers never feel lost, as they churn out a healthy backlog of clear, high-impact tasks driven by customer feedback. All these factors compound to make a meaningful difference for the users!
-
Shailesh Ahuja posted thisThe classic travelling salesman problem is considered extremely hard (NP). You are given a list of cities and the distances between them. Find a route to cover all cities in less than X miles. It's quite complicated to solve. However, verification of a given solution is far easier. Add up all the distances and check whether it's less than X. This is the fundamental reason AI significantly increases productivity. The task shifts from building to verification. For a seasoned dev, verifying X lines of code solve a problem is much easier than writing those X lines with no assistance.
-
Shailesh Ahuja shared thisWhat if your research assistant could fact-check itself in real-time? We just launched our chat agent at jenni.ai, and our users are already reporting 15-20% higher satisfaction scores within the first week. Compared to regular chat, an agent can utilise smart tools to delve deeper and provide more comprehensive answers over multiple steps. We lean heavily on backing all claims with citations and steer the model away from its internal knowledge. This reduces the usefulness for non-academic conversations, but it’s totally worth it for our audience, where the primary need is no hallucinations and facts based on published research. Excited to help researchers save time and boost their research confidence!
-
Shailesh Ahuja liked thisShailesh Ahuja liked thisMem0 migrated 400M+ agent memories from pgvector to turbopuffer, solving a semi-selective filtering problem that spiked tail latencies in Postgres → 150k+ isolated search indexes → 70ms p90 hybrid retrieval latency → 97% vector recall@10 https://tpuf.link/mem0
-
Shailesh Ahuja liked thisShailesh Ahuja liked thisNLB wants Singaporeans to read 15 minutes a day. The incentive: virtual coins, redeemable for cash. I understand the instinct, but I don’t think we actually know why people don't read-- no time, no habit, competing content, or something else? Those all require completely different fixes. A cash incentive only really works on one of them: it assumes the barrier is motivation. If the barrier is habit or opportunity instead, paying people to record 15 minutes of reading a day won't overcome that. There's also a risk in the other direction: motivational crowding out. When you pay someone for something that should be intrinsically rewarding, the behaviour disappears when the incentive does. It’s no longer "I read because I enjoy it" but "I read because I will be rewarded." The risk is that our measure of good becomes coin accumulation, not pleasure for or frequency of reading. One is easy to put on a dashboard; the other is harder to observe, track, influence. If I were designing this, I'd start somewhere slower and less trackable: one curriculum hour a week, every school (kindergarten to poly/ JC), for book-club style discussion-- not instruction, 𝘥𝘪𝘴𝘤𝘶𝘴𝘴𝘪𝘰𝘯. Peers talking about what they're reading-- graphic novels, fan-fiction, translated literature. Teachers modelling that adults read for pleasure too. No coins involved. Yes, it’s harder to measure; schools will have to give up instruction time; it’s slower to show results, but this is closer to how reading habits actually take root-- through social modelling, not microtransactions. https://lnkd.in/ge62HuuU
-
Shailesh Ahuja liked thisShailesh Ahuja liked thisThe next phase of AI runs on specialized intelligence — and we're building that in Singapore. Fireworks helps companies transform general-purpose models into specialized intelligence trained on their own data and deliver it at scale on the top-performing inference stack. The result is AI tailored to each business, often delivering frontier-quality performance at a fraction of the cost. We're standing up our first APAC team. We're looking for people who want to define what Fireworks becomes in the region. If you're strong in go-to-market or ML/infra engineering, take a look: Four founding roles 📍 Singapore — links in the first comment 👇 • AI Field Engineer — the technical bridge between our platform and enterprise customers @Shashank Verma • Enterprise Account Executive — owning our APAC go-to-market from the front @Sean Kelley • Applied Machine Learning Engineer • Member of Technical Staff, Cloud Infrastructure If those are the problems you want to work on — or you know the right person — let's talk. #FireworksAI #Hiring #Singapore #AIInfrastructure #MachineLearning #AI #TechJobsSG
-
Shailesh Ahuja liked thisShailesh Ahuja liked this🫡 I am looking for a partner-in-crime to build and scale Mistral in APAC 🔥 The mission is simple - hiring needs to mirror the breakneck speed at which our business is growing, across Singapore, Korea, India and Australia (for now). JD below, but fundamentally we're looking for: ✔️ First principles thinkers with a strategic mindset and execution focus ✔️ Jobs-to-be-done approach and a low ego mindset of humility and curiosity -- I could report to you or you could report to me, it should not matter ✔️ Ideally you've worked at high growth startups and have seen the 1-100 ✔️ Extremely collaborative in driving outcomes as a team, along side being a sparring partner to stakeholders ✔️ Creative problem-solvers who thrive in solving impossible problems - there is no easy here https://lnkd.in/ghfd6M4r
-
Shailesh Ahuja liked thisExcited to welcome Kinisi to Bear Robotics. I have spent much of my career thinking about how robots can operate reliably in the real world — not just in demos, but in complex, unstructured commercial environments. The next major step for robotics is Physical AI: systems that can perceive, reason, move, and interact with the world around them. Bear has built a strong foundation in autonomous mobility and real-world deployment. With Kinisi joining us, we are adding a critical missing piece: robotic manipulation. This combination gives Bear a much more complete Physical AI platform — mobility, manipulation, perception, autonomy, and deployment at scale. I’m very impressed by the Kinisi team. They bring deep technical capability, strong execution, and the kind of builder mindset that matters in robotics. Excited to welcome them to Bear and build the next chapter together.Shailesh Ahuja liked thisBear Robotics has signed a definitive agreement to acquire Kinisi Robotics — the team behind the KR1 humanoid. For nine years, we've built the world's most widely deployed fleet of autonomous service robots: more than 16,000 shipped into commercial service across hospitality, healthcare, retail, logistics, and real estate, all working as a single coordinated fleet through agentic multi-robot orchestration. But moving through the world is only half of physical work. The other half is handling it — picking, placing, sorting, manipulating the objects around us. That's exactly what Kinisi brings, and it completes our end-to-end Physical AI platform. What Kinisi adds: • The KR1 humanoid robot • Proprietary Vision-Language-Action (VLA) and Robot Foundation Model (RFM) — built on the NVIDIA stack, as a member of NVIDIA for Startups Inception • In-house gripper and end-effector design, plus a low-cost glove for capturing manipulation data • A world-class Physical AI team and a new European engineering hub in Bristol, UK And it brings someone home: Brennand Pierce, a Bear co-founder and Kinisi's founder, returns as our Chief Robotics Officer on closing. Most robotics companies are still working to get from a pilot to a product. We're coming from the other direction — already deployed at scale — and adding the one capability that turns a fleet into a platform for general-purpose Physical AI. See the KR1 live at Automate Show 2026 in Chicago, June 22–25. This is the start of a much bigger chapter. Read the full announcement here: https://lnkd.in/edf9NqRC #PhysicalAI #Robotics #HumanoidRobots #GeneralPurposeRobots #AI #Automation
-
Shailesh Ahuja liked thisShailesh Ahuja liked thisThe biggest treasure hunt happening in the AI world now is finding the hidden gem models.💎 A hidden gem model is a model that cost a fraction of the frontier models but delivers comparable performance for the right tasks. You can’t find these hidden gems through vibe checks. And the fact that expensive frontier LLMs like Fable may top leaderboards still does not answer the industry’s question around ROI once the excitement dies down. They can only be uncovered through rigorous AI evals. Evals show the relationship between cost spent and value delivered. More importantly, they help identify the point where a cheaper model has reached its maximum practical capability, and where paying for a frontier model begins to generate meaningful returns (the inflection point). As the market shifts toward usage-based pricing and bring-your-own-API-key business models, this research is becoming increasingly important, not just for legal AI builders, but also for legal teams choosing AI tools and designing AI workflows. This is one of the research questions we’re exploring at Legal Benchmarks 📊, together with a global community of legal practitioners and researchers. If this is a topic you’re interested in, we’d love to hear from you. 🚀
-
Shailesh Ahuja liked thisShailesh Ahuja liked thisLegora searches 2B+ legal documents on turbopuffer → strict per-matter data isolation with CMEK → 10x lower tail latency than Postgres → 98% avg recall@10 tpuf.link/legora
Experience & Education
-
Jenni
View Shailesh’s full experience
See their title, tenure and more.
or
Projects
-
TickerBot
See projectA conversational game on slack. Test your knowledge of the stock markets, challenge your colleagues in a friendly financial duel and find out who is the best investor in your team.
-
Sentiment Blade
See project• Created a sentiment analysis engine using Python (Information Retrieval and analysis)
• Objective: To find out the most effective method to classify reviews based on sentiments.
• Researched NLTK, WordNet, Stanford Parser, Thesaurus
• Wrote a crawler to extract reviews of restaurants from www.yelp.com
• Creating a dictionary based classifier, combining it with rules to improve precision and recall -
Computing Innovation and Design Programme (CIDP)
See project• CIDP emphasises the application of a complete software development cycle towards the construction of a web-based software system.
• Team of five people, developed a website serving as liaison between customers and companies to plan weddings.
• Used PHP with Yii framework to develop a user friendly and powerful website.
• Developed Budget algorithm, to assign budget to different products automatically.
-
Using a Genetic Algorithm to Find Parameters for Intelligent Agents in a Multi-agent System
-
– Designed a system of 2 communicating agents, which can achieve a high score in Tileworld by dropping tiles into
holes, while avoiding obstacles. A Genetic Algorithm was used to find the correct parameters for the agents.
– Technologies: Java, Mason (Multiagent simulation toolkit), Watchmaker Framework for Evolutionary Computation
– Individual Report: http://myexpressions.in/files/GA.pdf
– Presentation: http://youtu.be/6SzM8eR8Rrs, Code: https://github.com/shaileshahuja/TileWorldOther creators -
-
Tambola
-
Honors & Awards
-
Dean's List
Nanyang Technological University
Dean's List for academic year 2013-2014, awarded to those with exceptional academic performance
Test Scores
-
GRE
Score: 333
Languages
-
English
Full professional proficiency
-
Hindi
Native or bilingual proficiency
Organizations
-
Toastmasters International
Area Governor
- Present• Responsible for mentoring and guiding 5 clubs including UBS, Credit Suisse, NTU Toastmasters. • Organised Club Officer Training and gave a presentation on being a Vice President (PR) of a Club. • Past Vice President (Education) of NTU Toastmasters club. • Active member of the club, giving and evaluating speeches, being the host and organizing chapter meetings and contests. • Achieved Competent Leader (CL) and Competent Communicator (CC). • Gave an inspiring speech at the welcome tea, to…
• Responsible for mentoring and guiding 5 clubs including UBS, Credit Suisse, NTU Toastmasters. • Organised Club Officer Training and gave a presentation on being a Vice President (PR) of a Club. • Past Vice President (Education) of NTU Toastmasters club. • Active member of the club, giving and evaluating speeches, being the host and organizing chapter meetings and contests. • Achieved Competent Leader (CL) and Competent Communicator (CC). • Gave an inspiring speech at the welcome tea, to attract new members.
Recommendations received
2 people have recommended Shailesh
Join now to viewOther similar profiles
Explore more posts
-
Vedran B.
Just to make clear, if anyone else wants to ask: A tool that I built for myself, and have been using successfully, called NpgsqlRest IS NOT an in-place replacement for PostgrREST/Supabase. It is magnitude-level better and more capable, but you can't just replace PostgrREST/Supabase with NpgsqlRest and expect it to work. Here are the main differences: 1) Direct Table/View Access. I hate this. I think this is an antipattern; to be honest, it brings nothing but trouble. I prefer all access to go through routines (functions/procedures). And then you can choose to expose them as a REST API and configure them as you see fit. Routines are the primary level of abstraction, and hard-coupling tables directly to the client is not a good idea. Yes, you heard me right. Nothing new, I've been saying this for years. If you want to do a quick CRUD API over some tables, you can always ask your AI friend to generate. And then later, you can tweak details as you see fit, or even refactor the entire schema, once you have those client contracts enforced. You can't do that with those dynamic queries that PostgrREST/Supabase generates for you. I added just basic CRUD table-access endpoints, but PostgrREST/Supabase is much more advanced in this area. Instead of catching up, I am closer to removing it completely. 2) Per endpoint configuration PostgrREST/Supabase, you must configure all or nothing, as far as I know. With NpgsqlRest, you configure just what you want and how you want it. 3) Authorization PostgrREST/Supabase is using Row-Level-Security to tweak authorization levels and to let you set which tables a user has access to, for example: create policy "Users can view own data" on users for select using (auth.uid() = id); This is insane. Not only can you not set authorization per endpoint, but it forces you to go to the table level, and then RLS forces you to inject a hidden filter in ALL of your queries. Good luck tweaking indexes and partitions after that. In contrast, NpgsqlRest has a far more elegant, role-based approach with SQL comment annotations. You can't get simpler than that. comment on function api.get_user(int) is ' HTTP GET @authorize admin, user '; And, no, I don't have plans to reach feature parity with PostgrREST/Supabase. Not even if someone paid me to do it. PostgrREST/Supabase should reach feature parity with NpgsqlRest. https://lnkd.in/gcSSSRuH
4 Comments -
Princy Melchius
There’s a moment every professional reaches when they ask themselves “Am I still enjoying this… or am I just good at it?” Sai Preethi Polu reached that moment too. She was already a developer. Comfortable with code. Doing what many would call the “safe” thing. But she was curious about something else: how complex systems are explained, not just built. That curiosity brought her to Authoring Minds. She didn’t jump blindly. She learned the fundamentals. Then went deeper —> DITA, Oxygen XML, structured documentation. She asked questions. A lot of them. And she practiced until clarity became habit. I trained her through this transition, and what stood out wasn’t speed; it was intent. She wasn’t trying to escape development. She was trying to translate it. Today, she works as a Technical Writer with a Fortune 500 company. But the real win isn’t the brand name. It’s the confidence to shift paths without starting from zero. It’s proof that technical knowledge doesn’t disappear; it evolves. Sai Preethi’s journey is a quiet reminder: "Careers don’t have to restart. Sometimes, they just need a new direction." And when the foundation is right, the transition becomes possible. #Technicaldocumentation #casestudy #Technicalwriting
2 Comments
Explore collaborative articles
We’re unlocking community knowledge in a new way. Experts add insights directly into each article, started with the help of AI.
Explore MoreOthers named Shailesh Ahuja
3 others named Shailesh Ahuja are on LinkedIn
See others named Shailesh Ahuja