--- title: "Subandhu - - Machine Learning Engineer in Bengaluru | Tal" description: "Meet Subandhu -, Generative AI & ML Engineer | Agentic AI, LLMOps, RAG, Multi-Agent Systems | Building Production-Ready AI Products based in Bengaluru. AI stack includes claude code, codex, and cursor. Explore public work links including github." image: "https://tal.club/meet/subandhu/share-image" --- # Subandhu \- on Tal ## Headline Generative AI & ML Engineer \| Agentic AI, LLMOps, RAG, Multi\-Agent Systems \| Building Production\-Ready AI Products ## About I help organisations translate cutting‑edge research in generative AI, agentic systems and Retrieval‑Augmented Generation \(RAG\) into reliable, scalable products that drive revenue, reduce costs and delight users\.As a Generative AI / ML Engineer with over 3 years of experience across fintech, gaming and enterprise software, I specialise in transforming advanced AI research into production‑grade systems\. My work at Gameopedia demonstrates this: I productionised a video‑understanding agent using Gemini 2\.5 Flash, Agno AI and LangGraph, processing 3,000\+ videos and increasing watch‑time by 1\.2%\. I also built a sentiment‑analysis pipeline that ingested 20k\+ reviews, reduced costs 26×, and delivered insights across 60\+ categories\.Previously, at a stealth fintech startup, I designed hybrid RAG systems combining Graph‑based retrieval and vector databases to ingest 1M financial prospectuses, enabling an intelligent equity research agent\. At Arcesium \(DE Shaw\), I translated 350\+ Perl functions to Python, fine‑tuned LLMs for test‑case generation \(24k unique cases\), and built interactive knowledge graphs for coding agents\.My core stack includes Python, PyTorch, TensorFlow, LangChain, Hugging Face Transformers, LangGraph and Kubernetes\. I’m comfortable with cloud services \(AWS, GCP\), MLOps \(Docker, Kubernetes, Kubeflow, MLflow\) and vector databases \(Pinecone, Milvus\)\. Colleagues describe me as a technical generalist with product sense—I thrive on cross‑functional collaboration, clear communication and delivering measurable impact\. ## Location Bengaluru ## AI stack claude code codex cursor hugging face OpenAI Agents SDK Qdrant ## Work experience Machine Learning Engineer at Pocket FM, Jun 2026 \- Present · 2 mo, Bengaluru AI Engineer at Gameopedia, Dec 2024 \- May 2026 · 1 yr 6 mos, Hyderabad, Telangana, India, Built and productionized agentic video understanding pipelines for long\-form content using Gemini 2\.5 Flash, enabling automatic highlight and short\-form extraction from gaming streams, sports, trailers, vlogs, and 10\+ hour videos; processed 3,000\+ videos and generated 50\+ unique highlights per title, increasing watch\-time by 1\.2%\.  • Designed an end\-to\-end multimodal sentiment intelligence system using OpenAI Agents \+ Batch API, ingesting 20,000\+ Steam and Reddit reviews \(text \+ images\), mapping insights across 60\+ taxonomy categories for game marketing and publishing teams; reduced inference cost from $8/game to $0\.30/game and deployed the system as an MCP server\.  • Developed an agentic taxonomy generation and classification framework by fine\-tuning Small\-LM \(360M, LoRA\), DeepSeek\-R1 \(Ollama\), and GPT\-3\.5, orchestrated via LangGraph; achieved 78% semantic similarity and unlocked 100\+ new structured game insights for downstream analytics and discovery\.  • Architected an auto\-evolving prompt refinement and context\-engineering pipeline using Gemini 2\.5 Flash and OpenAI Responses API, enabling self\-improving classification workflows; delivered 150\+ auto\-enhanced definitions and improved classification accuracy from 33% → 88% within three refinement iterations, validated across 60\+ controlled experiments\.  • Worked at the intersection of research → production, translating GenAI experimentation into scalable, cost\-efficient systems, with a strong focus on agent orchestration, observability, and business\-aligned metrics \(accuracy, cost, engagement lift\)\. Machine Learning Engineer at Stealth Startup \(Fintech\), Aug 2024 \- Nov 2024 · 4 mos, • Performed EDA and data quality audits on 1M\+ financial prospectuses; designed hybrid RAG \(Graph \+ LanceDB\) with chunking strategies, embedding model selection, and retrieval benchmarking; deployed intelligent financial analysis agent using AWS Bedrock, Groq, and LangGraph\. • Built and evaluated ML classifiers for financial document categorization; compared embedding\-based approaches \(cosine similarity, ANN search\) vs\. fine\-tuned classifiers for retrieval accuracy; implemented POCs for HYDE RAG, CRAG, and Reflexion Agents\. • Implemented Equity Research Agent serving 7\+ analyst functionalities with Human\-in\-the\-Loop oversight and structured financial reasoning\. Gen AI Engineer at Stealth AI Startup, Aug 2024 \- Nov 2024 · 4 mos, Remote, Primary Focus: Creating research based POCs for conversational AI 1\. Implemented graphRAG \(using neo4j and microsoft\-graphRAG\) to analyse multiple documents for chatbot and insight extraction\. 2\. Created structure design for Agentic framework targeting research report generation\. 3\. Worked with parallelization of RAG chains with langchain reduce generation time\. 4\. Setup chat window history to maintain chat context over longer conversations\. 5\. Created langgraph based ReAct agents targetted to specific report component generation tasks\. Deploying 9 agents working in parallel under Supervision to generate research document pdfs\. 6\. Created evaluation and implementation notebooks around various RAG techniques\(CRAG, Agentic RAG, Hyde\)\. Implemented hybrid\-search with LanceDB, using custom embeddings modules\. 7\. Researching Anthropic’s Contextual Retrieval to assist generation results\. 8\. Worked with query re\-rankers: as agents and with Colbert, Cohere and OpenAI reranker to demonstrate improvement in document retrieval for chatting and end pdf\(research report\) creation\. 9\. Extensively worked with ReAct agents and function calling capabilities of LLMs to improve structured data extraction from documents\. 10\. Experimented LlamaParse for chatting with Ms Excel workbooks across multiple sheets, maintaining context during extraction and question answering\. Machine Learning Engineer at Arcesium, Aug 2023 \- Aug 2024 · 1 yr 1 mo, Hyderabad, Telangana, India, Language model–based code translator: Orchestrated automated migration of 350\+ Perl functions to Python using Vicuna 13b, achieving 75% functional accuracy; developed an interactive UI for real‑time translation and evaluated 15\+ LLMs \(LLaMA, GALACTICA, etc\.\)\. Tax character prediction: Developed and hosted an image classifier that processed 250\+ documents and achieved 98% accuracy; performed data exploration on 1M\+ rows to identify trends in straddle tax character\. LLM‑based test case generation: Designed an automated framework using Llama2, Mistral7b and CodeLlama to generate 24k unique test cases, improving test coverage from 32% to 80%\. Domain‑aware coding assistant: Built knowledge graphs \(depth 7\) for LLM‑driven coding assistants to enable end‑user exploration, integrated AWS KnowledgeBases via vectorization and synchronisation, and implemented LangFuse for observability on 10k\+ LangChain flows\. Token cost & data safety improvements: Integrated Nemo Guardrails and anonymisation into 50\+ LangChain flows, reducing token usage costs by $3 per 10 LLM generations and preventing proprietary data leakage\. Machine Learning Engineer at Arcesium, Mar 2023 \- May 2023 · 3 mos, Bangalore Urban, Karnataka, India, Evaluated 15\+ code translation approaches ranging from Heuristic to Generative AI\. Evaluated entire space of open source LLMs\. Setup pipeline for automated translation of legacy Perl code to Python optimised for functional accuracy\. ## Education Technology Leaders Program, Machine Learning, Plaksha University Bachelor of Technology \- BTech, Computer Science, KIIT \- Kalinga Institute of Industrial Technology ## Public links - github: https://github\.com/agravans [View the original profile](https://tal.club/meet/subandhu) ```json { "@context": "https://schema.org", "@graph": [ { "@type": "Person", "@id": "https://tal.club/meet/subandhu#person", "name": "Subandhu -", "url": "https://tal.club/meet/subandhu", "image": "https://tal.club/meet/subandhu/profile-photo.jpg", "description": "Meet Subandhu -, Generative AI & ML Engineer | Agentic AI, LLMOps, RAG, Multi-Agent Systems | Building Production-Ready AI Products based in Bengaluru. AI stack includes claude code, codex, and cursor. Explore public work links including github.", "jobTitle": "Machine Learning Engineer", "worksFor": { "@type": "Organization", "name": "Pocket FM" }, "hasOccupation": { "@type": "Occupation", "name": "Machine Learning Engineer" }, "homeLocation": { "@type": "Place", "name": "Bengaluru" }, "alumniOf": [ { "@type": "EducationalOrganization", "name": "Plaksha University" }, { "@type": "EducationalOrganization", "name": "KIIT - Kalinga Institute of Industrial Technology" } ], "knowsAbout": [ "claude code", "codex", "cursor", "hugging face", "OpenAI Agents SDK", "Qdrant" ], "sameAs": [ "https://github.com/agravans" ], "mainEntityOfPage": { "@id": "https://tal.club/meet/subandhu#webpage" } }, { "@type": "ProfilePage", "@id": "https://tal.club/meet/subandhu#webpage", "url": "https://tal.club/meet/subandhu", "name": "Subandhu - on Tal", "description": "Meet Subandhu -, Generative AI & ML Engineer | Agentic AI, LLMOps, RAG, Multi-Agent Systems | Building Production-Ready AI Products based in Bengaluru. AI stack includes claude code, codex, and cursor. Explore public work links including github.", "dateModified": "2026-08-09T04:51:50.791Z", "isPartOf": { "@id": "https://tal.club/#website" }, "mainEntity": { "@id": "https://tal.club/meet/subandhu#person" }, "primaryImageOfPage": { "@type": "ImageObject", "url": "https://tal.club/meet/subandhu/profile-photo.jpg", "contentUrl": "https://tal.club/meet/subandhu/profile-photo.jpg", "caption": "Subandhu - profile photo on Tal" }, "inLanguage": "en-IN" } ] } ```