About
Have close to 10 years experience of entire Data science use case lifecycle, starting…
Articles by Subhojit
-
Theano on GPU for Windows
Theano on GPU for Windows
Python has many efficient Deep Learning libraries. Max of them runs on Linux machine efficiently.
2 Comments -
First try with Web Scraper Program
First try with Web Scraper Program
Use case was like following: 1. Input will be a topic name and number of expert to find like: cryptography 2…
1 Comment -
Clustering algorithms Benchmarking
Clustering algorithms Benchmarking
clValid package in R..
2 Comments
Activity
11K followers
-
Subhojit Mukherjee reposted thisSubhojit Mukherjee reposted thisOne transformer just replaced 15+ candidate generators, pre-ranking, and ranking models - and beat them on live traffic. The team at Yandex Music published a technical report on Sona, a single-model generative recommender that replaced their entire production cascade for My Vibe on smart speakers. The results from the online A/B test: +4.53% Active Users (the primary metric), +6.30% Total Listening Time, and +11.42% Likes. That Active Users uplift is 2.35x what Argus, their strongest previous model, delivered on the same surface. How it works under the hood: A shared encoder processes the user's chronological engagement history - plays, skips, likes, dislikes - once per request. Two modules consume that representation: an autoregressive decoder and a Ranking Module. So candidate generation and ranking run inside one served model, paying the encoder cost a single time. Instead of predicting raw track IDs, the decoder generates Semantic IDs: each track is tokenized into a 3-code tuple drawn from three 32,000-entry codebooks. The tokenizer pipeline is clever - a frozen multimodal LLM (Qwen2.5-Omni) encodes the first 90 seconds of audio plus metadata, a refinement transformer trained on collaborative listening pairs aligns the embeddings with actual user behavior, and residual K-means quantizes them. Constrained beam search (width 1024) then generates tuples, which map back to catalog tracks. Ranking supervision comes via distillation. A 0.6B-parameter Teacher Ranker - pre-trained with next-item prediction on a full year of engagement logs, then fine-tuned for multi-head ranking - scores the decoder's own beam-search rollouts plus logged impressions, and the lightweight Ranking Module learns to reproduce those scores. The teacher never runs at serving time. Two more details worth noting: History Compression lets the model attend to 8,192 events by applying the deep layers only to the most recent 2,048, roughly halving inference cost. And the whole system trains continuously online - a user event reaches an updated production model in about 45 minutes at the median, with fresh weights shipped every 10 minutes. No hand-engineered features anywhere. Just logged events, learned representations, and scale. The era of multi-stage recommendation cascades may be ending. This post is in free😅 support of friends at Source Strong AI (Do check them out)
-
Subhojit Mukherjee shared thisGood oneSubhojit Mukherjee shared thisA Google Distinguished Engineer just open-sourced how Google builds production agents. 291 pages. 22 chapters. All code public. This is the playbook Google engineers use internally. "Building AI Agents: From Design Patterns to Production" Written by Antonio Gulli (Sr Director, CTO Office) and Anant Nawalgaria (Senior Staff ML Architect). The core loop: Perceive → Plan → Act → Observe. 4 patterns implemented end-to-end: → ReAct → Chain-of-Thought → Reflection → Plan-and-Execute Each chapter builds Atlas, a research assistant that evolves from a 50-line script into a self-correcting multi-agent system. 100% of author royalties go to Save the Children. Learn Agent Engineering → Support Kids → https://lnkd.in/eMQEVVEk
-
Subhojit Mukherjee shared thisThis is gemSubhojit Mukherjee shared thisWhat happens when you ask AI research agents to do the job of a financial analyst? Google Cloud AI Research and UCLA tested 17 systems on analyst-grade financial research questions. The result? Even the best systems struggled. The researchers introduced: → FinanceGym: a benchmark for testing financial research agents → FinanceHarness: the system around the LLM that helps it research, use tools, analyse information, and produce its final report The harness is everything surrounding the model: its tools, instructions, and workflow. FinanceGym contains 400 questions about real companies, designed to reflect equity research tasks such as analysing a company, its competitors, and what could happen next. 👉 How did they test the agents? Each question has a specific cut-off date. The agent can search more than 100 million dated articles, but cannot access anything published after that date. The evaluation then separates two things: → What could the agent establish using information available at the time? → Could it correctly anticipate what happened afterwards? And this is where the results become interesting. 👉 Finding the evidence is much easier than making the right call. FinanceHarness scored: → 45.7% on information available before the cut-off → only 11.8% on outcomes that became known afterwards The same pattern appeared across all 17 systems. In other words, these agents are much better at researching what is already known than reasoning about what comes next. 👉 Another result caught my attention: the harness itself made a big difference. Using the same Qwen3.6-27B model and the same search tool: → Search only: 25.3% → Full FinanceHarness: 32.4% → FinanceHarness + reinforcement learning: 32.8% So the specialised agent architecture added about 7 percentage points. Reinforcement learning added only another 0.4. A useful reminder that building better financial agents is not only about improving the underlying model. The system around it matters too: tools, workflow, context, and instructions. 👉 Some limitations to keep in mind: Gemini-3.5-Flash generated the questions and checklists and also graded the answers, with no reported human-judge agreement. 82% of questions passed expert review. The corpus and checklists are not public, and the data is limited to 2025 English web articles, with no filings or paid data. 👉 My 3 takeaways: - The underlying model still matters a lot. - The system around the model can significantly improve performance, even without changing the model itself. - Forward-looking judgment remains the biggest gap. Better retrieval helps an agent understand what is known, but does not automatically make it good at predicting what comes next. And for me, that last point is probably the most important one for financial research agents. Finding the right information and making the right investment call are two very different problems. Links to the GitHub repo and paper👇 Enjoy this? ♻️ Repost and share.
-
Subhojit Mukherjee shared thisMost agent architectures share a hidden assumption: Every decision inside the loop goes through a generative LLM. That assumption is worth questioning. —— Came across JEV recently, and the core idea cuts deeper than it first appears. Instead of generating text token by token, Jev takes a state and answers structured decision questions directly — with probabilities and confidence scores. It is described as a "System One" model: fast, structured decisions, sitting inside the agent loop. That reframes the architecture problem. —— A typical agent loop today: User → LLM → Tool → LLM → Tool → LLM → Final answer A more differentiated architecture: User ↓ Reasoning LLM ↓ Decision Layer ↓ Tool / Action ↓ Decision Layer ↓ Reasoning LLM (only when depth is actually required) The difference isn't just a performance optimization. It's about using the right model for the right kind of decision. —— A large model should own the expensive parts: • Reasoning and planning • Synthesis • Open-ended generation A smaller, faster decision model handles the repetitive ones: • Which tool should I use? • Is this request urgent? • Should this action be allowed? • Does this need escalation? • Which model should handle this subtask? —— This is where Jev adds a useful lens. Jev is an independent implementation that explores direct probability decisions — without AUTOREGRESSIVE generation. It exposes Choice, Noul, and Score decisions where applications can consume the model's output directly. No generation step. No parsing step. —— Maybe the future of production agents isn't one large model doing everything. It's a coordinated system of specialized models: 🧠 Reasoning model — System 2 thinking ⚡ Decision model — System 1 routing 🔎 Retrieval model — knowledge access 🛡️ Safety / policy model — constraint enforcement 🎯 Ranking model — candidate selection 🛠️ Action layer — execution Large model handles depth. Decision layer handles velocity. —— At scale — when an agent makes hundreds of routing decisions during a single long-running workflow — that separation has real consequences for latency, cost, and reliability. LangChain is already experimenting with this pattern for model routing and tool-call guardrails. —— Performance Caveat: The speed and cost claims around Jev are vendor claims. Open source variant Open-Jev explicitly does not claim to reproduce TypeSafe's proprietary RLCD approach or the advertised speedups. Quality, calibration, and latency still need independent evaluation. —— But the architectural question is worth looking into : What if "one LLM for every decision" is just the first draft of agent engineering? The next generation of agent systems may be less about finding a model that can do everything, and more about designing the right combination of models for different kinds of decisions. #Jev #AI #GenerativeAI #AIAgents #LLM #MachineLearning #AgentArchitecture
-
Subhojit Mukherjee shared thisGood approach.Subhojit Mukherjee shared thisYour AI Agent doesn't need a bigger ontology. It needs one that updates itself. Here's the trap most teams fall into. You hand-build a semantic layer, ship it, and it's a snapshot from day one. Data shifts, workloads shift, and the definitions stop matching how the work is actually done. Nobody has time to keep rewriting it, so agents go back to inferring meaning on every task and paying for it in tokens. A paper on EvoOntology makes the case for a self-evolving alternative: an ontology the agent trains as it works, not a document you maintain. 📌 What makes it self-evolving comes down to three things. 1. It learns from the real work The layer is built from the actual workload and checked against the source data, then keeps adapting as agents use it. It tracks reality instead of drifting from it. 2. It changes only when the change wins Updates aren't trusted blindly. A new version ships only if it beats the previous one on the same task and budget, and every change stays inspectable and reversible. 3. It grows without bloating the prompt Agents pull only the slice of meaning a step needs, so the ontology can keep expanding without dragging more context, and more cost, into every call. The shift underneath: your semantic layer stops being documentation you maintain and becomes a state that sharpens as the system runs. That is the version of ontologies I want under my agents. 👇 Would you trust an agent to evolve your ontology on its own, or does that stay a human job? 📌 More AI agent insights → https://lnkd.in/gatGAt9X 📘 Grab your copy → The Human-Agent Orchestrator: https://amzn.to/4gEIzmv Save 💾 React 👍 Share ♻️ - follow for daily AI agent breakdowns #AIAgents #AgenticAI #GenerativeAI #EnterpriseAI #Ontology #DataAgents #evoontology
-
Subhojit Mukherjee shared thisGood articleSubhojit Mukherjee shared thisNo course will teach you Harness engineering. No tutorial will either. Two GitHub repos will. → learn-harness-engineering ↳ Walks through every core mechanism ↳ Understand the internals before building on top ↳ https://lnkd.in/gq3mF-8Y → pi (earendil-works) ↳ Module design and implementation, end to end ↳ When something is unclear, let AI explain the logic ↳ https://lnkd.in/gvHqstzq Two more repos worth knowing: → harnesscode ↳ Long-running agents based on the Anthropic Harness paper ↳ External memory + small-step execution + auto-loop ↳ https://lnkd.in/gsnQ9D3t → llm-wiki-compiler ↳ Structured knowledge that compounds over time ↳ Outer loop handles decisions, inner loop handles knowledge ↳ https://lnkd.in/g7DqFPZv The method is simple. Read real implementations. Build something small. Use AI as your tutor when you get stuck. Not docs. Not tutorials. Source code. I'm Shrey Shah & I teach AI assisted coding and agents.
-
Subhojit Mukherjee reposted thisSubhojit Mukherjee reposted thisBuild AI agents on a time-aware knowledge graph! Utopia is an open-source knowledge system that turns documents, databases, and connected sources into a temporal graph your agents can reason over. Most RAG systems are optimized for one question: what is relevant right now? That works until the underlying knowledge changes. A customer contract gets updated. A project owner changes. A policy is revised. A previous fact may no longer be true, but simply overwriting it means the system loses the history behind that change. Utopia handles this with a bitemporal knowledge graph. Each fact can track both when it was true in the real world and when the system learned about it. When something changes, the old fact is preserved instead of silently disappearing. That means an agent can reason about questions like: • What is true now? • What was true three months ago? • When did this information change? • What evidence was the conclusion based on? The graph is also ontology-aware, so documents are represented as entities, facts, and relationships instead of only chunks and embeddings. That gives the system more structure for reasoning across relationships, resolving entities, detecting conflicting facts, and deriving new information through explicit rules. Key capabilities: • Bitemporal knowledge graph for tracking how facts change over time • Provenance on facts so agents can trace where information came from • Conflict detection instead of silently overwriting contradictory information • Ontology-based reasoning across entities, relationships, and derived facts • Hybrid retrieval across full-text search, vector search, and graph traversal • MCP and agentic RAG support for exposing the knowledge layer directly to agents The interesting part is that this turns the knowledge base into more than a retrieval system. Instead of only finding relevant information, an agent can reason over what changed, what is still valid, how facts are connected, and where each conclusion came from. 100% open source. I've shared the GitHub repo in the comments!
-
Subhojit Mukherjee reposted thisSubhojit Mukherjee reposted thisYour recommender may be deleting the right answer before it ever gets ranked. A new study from researchers at the University of Illinois Urbana-Champaign takes apart Semantic IDs (SIDs), the backbone of modern generative recommendation, and finds a structural conflict hiding in plain sight. Quick refresher on how these systems work under the hood: an encoder (Sentence-T5, BGE, Qwen) turns item content into a continuous embedding, a quantizer (RQ-VAE, RQ-KMeans, OPQ/PQ, and others) compresses it into a short sequence of discrete tokens, and an autoregressive decoder like TIGER generates that token sequence with constrained beam search. The catch: every generated token is also a filtering step. Discard a prefix, and every item under it is permanently gone, before final ranking even begins. What the paper found across three Amazon domains and eight SID constructors: - SID neighborhoods recover only 32.2% of the encoder's ten nearest neighbors on average. Coarse category structure survives; fine local structure does not. - Simply reordering an item's description fields (same words, same facts) changes 38.4% of exact SIDs on average, even though the encoder still retrieves the correct item first in 99.57% of controlled cases. Exact SIDs are not determined by meaning alone. - The consequence at decoding time: after the final semantic token, TIGER retains only 35.4% of Office targets and 24.4% of Scientific targets that an independent item scorer had ranked in its top 10. The model's own tokenization prunes items a separate model considers highly relevant. Their fix, Item-Supported Decoding (ISD), is refreshingly pragmatic. No retraining, no new parameters. Before generation, build one ranked item set from the user's history using any ranker: simple training statistics (item transitions plus co-consumption), SASRec, or UniSRec. During beam search, each candidate prefix gets two ranks: one from the decoder's sequence probability and one from the best-ranked item still living under that prefix. Reciprocal rank fusion merges them, so a prefix can survive either because the decoder likes it or because it shelters a highly relevant item. After generation, the same ranking reorders the output. Asymptotic decoding complexity stays unchanged. Result: NDCG@10 improves in every evaluated setting across TIGER and LIGER, with gains up to 31.2%. The takeaway is bigger than the method: SIDs are good at organizing items, but their fine token boundaries should not be the sole authority on which items survive generation. Identification and selection are different jobs. This post is in free😅 support of friends at Source Strong AI (Do check them out)
-
Subhojit Mukherjee reposted thisSubhojit Mukherjee reposted thisCan generative recommendation actually reach cold items? A new study from researchers at Renmin University of China and Alibaba Group says: only within narrow limits. Semantic-ID (SID) based generative recommenders like TIGER represent each item as a sequence of discrete semantic tokens (typically via RQ-VAE quantization of content embeddings) and predict the next item by autoregressively generating its SID with beam search. The promise: token recombination should let the model compose SIDs for items it never saw as training targets. The catch: most benchmarks use leave-one-out splits, where a "new" item for one user may already sit in another user's training history. This paper replaces that with an absolute-time sliding-window protocol - train strictly on the past, test strictly on the future. Under this setting, cold items make up to 95% of test targets, and unseen-item Recall@20 collapses to near zero. The token-level diagnosis is the interesting part: - Coldness is not binary. Cold items split into all-token-seen, any-token-unseen, and prefix-seen categories. Surprisingly, even items whose every token was observed in training remain nearly unreachable. - The reason is path dependency. In RQ-VAE, each later code encodes the residual left by earlier codes - so a SID is an ordered residual path, not a bag of reusable tokens. New combinations of seen tokens often land on unsupported paths. - Oracle-prefix probing (forcing the correct first one or two tokens) recovers many seen items but barely helps unseen ones. The bottleneck isn't coarse bucket selection - it's fine-grained path completion. - Comparing TIGER's token distributions with SASRec's bucket-aggregated scores shows SID decoding behaves like hierarchical semantic bucketing: early tokens pick coarse semantic regions, later tokens refine item-specific paths. Three diagnostic variants probe fixes: a factorized OPQ tokenizer to reduce path dependency, a scorer interface that swaps exact decoding for candidate ranking (largest unseen-item gains), and cross-attention over dynamic edge/text context (strongest on evolving, text-rich datasets). Bottom line: SID generation is compositional, not open-ended. Reachability is bounded by learned token vocabularies and path support - a crucial distinction for anyone deploying generative retrieval on evolving catalogs.
Subhojit Mukherjee commented on a post
2y
Congratulations
-
Subhojit Mukherjee liked thisSubhojit Mukherjee liked thisMust-read paper from Google. If your agents run on SKILL.md files, save this one: A 9B model with evolved skills beat a 27B model without them: 47.4% vs 39.4% average across 5 benchmarks. Same family (Qwen), no fine-tuning. Most skills are written by hand. Skill evolution lets the agent write its own: run tasks, read traces, patch the skill, keep it if validation improves. Current methods (EvoSkill, SkillOpt, Trace2Skill) scatter what they learn across optimizer history. The paper proposes a method called WikiSkill. It adds a persistent wiki between raw traces and skills, inspired by Karpathy's LLM Wiki. After each run, a maintainer agent compiles the traces into failure and success patterns. A proposer agent reads the wiki to patch SKILL.md files. Bad edits roll back, but the wiki never resets. ▫️ It beats prior methods. Its average score beat the best existing method by 3.3 to 12.0 points on all 5 models. ▫️ Bigger models gain more. Versus no skills, it added 12.3 points to Qwen 4B, 17.5 to 9B, and 23.9 to 27B. ▫️ Skills transfer between models. Qwen-9B scored 24.3% on SpreadsheetBench with no skills and 50.5% with skills written by Qwen-27B. ▫️ The wiki drives most of the gain. Without it, Gemini 3.5 Flash fell from 63.7% to 48.7%. A bigger model isn't the only upgrade path. Before paying for one, test what evolved skills do for the model you already have. Link to the paper in the comments. -- P.S. We’re building the best AI English tutor in the world. Try it to improve your speaking - it’s 15× cheaper than a human one → GetFluently.app
-
Subhojit Mukherjee liked thisSubhojit Mukherjee liked thisThis is what I'd read over the weekend if I wanted real examples on how top companies are (actually) using AI across ML and analytics. 👩💻 Personalization & Recommendations (ML) Foundation Model for Personalized Recommendation by Netflix 🔗 https://lnkd.in/dbJ-MJkV Recommending Travel Destinations by Airbnb 🔗 https://lnkd.in/dJvbWwVY Embedding-Based Retrieval for Airbnb Search by Airbnb 🔗 https://lnkd.in/dFvTZPUg Ads Candidate Generation using Behavioral Sequence Modeling by Pinterest 🔗 https://lnkd.in/d_5ikYhP 🤖 Agentic Analytics Scaling Global Storytelling: Modernizing Localization Analytics by Netflix 🔗 https://lnkd.in/d25WGhFT Why We Built "BlaBlaCar Data Copilot": Shifting Data Analysis Left by BlaBlaCar 🔗 https://lnkd.in/dW7wGjyK Dashboards Were Never the Destination by Hex 🔗 https://lnkd.in/dWqvz9hp Inside OpenAI's In-House Data Agent by OpenAI 🔗 https://lnkd.in/dEVizedA Our Multi-Agent Architecture for Smarter Advertising by Spotify 🔗 https://lnkd.in/dFexYGVA Unified Context-Intent Embeddings for Scalable Text-to-SQL by Pinterest 🔗 https://lnkd.in/dQZRypsW -- 📌 By the way, if you are a data scientist who wants to leverage AI in your day-to-day, checkout my free workshop series (starting September 24): https://lnkd.in/diN3Xhqx
Experience & Education
Projects
-
Location ID- Client project of American Express
From the Transaction submission data, we have to make clusters of locations and generate location Ids to identify business locations uniquely.
-
Cyber Forensic Analysis and Development
Analysis and research of certain Cyber Forensic issues were there. Two optimal applications had to be developed for given issues.
• First Issue was about System’s Clock Tampering. Due to malware or Human interference, System time can be tampered. As a result, Antivirus protection, SSL certificates and lots of Securities can be failed. So one application, as per CDAC requirements had to be developed.
• Second issue was with Browsing History in a machine filtering. If a machine is…Analysis and research of certain Cyber Forensic issues were there. Two optimal applications had to be developed for given issues.
• First Issue was about System’s Clock Tampering. Due to malware or Human interference, System time can be tampered. As a result, Antivirus protection, SSL certificates and lots of Securities can be failed. So one application, as per CDAC requirements had to be developed.
• Second issue was with Browsing History in a machine filtering. If a machine is running for 3+ years and more, there can be lakhs of web history generated. If any Computer crime is found in any Systems, then Cyber Forensic Investigator will be needing some efficient application to find some vile websites from the history. There was no application at that moment which can efficiently handle that situation in very short time. Required websites with the frequency were to be produced from web history.
These two solutions were developed using C# (Visual Studio 2013)
Honors & Awards
-
1st Dan Black Belt at Wado Kai Karata Do,India
JKF and AIKF
Languages
-
English
-
Organizations
-
College Photography Association
Editor
-
Recommendations received
6 people have recommended Subhojit
Join now to viewOther similar profiles
-
Rachana Baldania
Rachana Baldania
Diligent, results-driven Data Scientist with 9 years of experience delivering high-impact AI solutions across pharma, finance, audit, BI, e-commerce, and marketing domains. Proven expertise in AI, ML, Deep Learning, Generative AI, NLP/NLG, and advanced statistical modeling, with a strong track record of building scalable, production-grade systems.<br><br>Currently working remotely for Pfizer (USA) as a Data Scientist contractor through RandomTrees, India. Led and delivered 60+ data science projects for global clients including Kimberly-Clark, IBM–Pfizer, Nissan, Spar, Deloitte, Bayer, Microsoft, Reliance, Zomato, Maruti Suzuki, and Bharti Airtel.<br><br>MTech in Computer Engineering with specialization in NLP from Dharmsinh Desai University; secured 28th rank in Gujarat PGCET (38K+ candidates). Completed 20+ certifications and published 3 research papers in Data Science.<br><br>LinkedIn: https://www.linkedin.com/in/rachana-baldania-0435827b/<br><br>GitHub: https://github.com/Rachana-Baldania<br><br>ResearchGate: https://www.researchgate.net/profile/Rachana-Baldania/amp<br><br>Connect at: rbaldania2103@gmail.com<br><br>Thank you 🙋♀️
Explore more posts
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 More