Skip to content
View findsamirks-commits's full-sized avatar

Block or report findsamirks-commits

Block user

Prevent this user from interacting with your repositories and sending you notifications. Learn more about blocking users.

You must be logged in to block users.

Content in all repositories owned by your account will be closed.
Maximum 250 characters. Please don’t include any personal information such as legal names or email addresses. This note will only be visible to you.
Report abuse

Contact GitHub support about this user’s behavior. Learn more about reporting abuse.

Report abuse
findsamirks-commits/README.md

πŸ‘‹ Hi, I'm Samir Sandhwar

Senior Retail Professional & Merchandising Head | Agentic AI & Data Science Engineer | Business Head
Bridging 24+ years of high-volume retail P&L leadership with production autonomous multi-agent AI architectures.

LinkedIn Email GitHub List Hugging Face Streamlit Render

🧠 Executive Profile & Foundations

I bring over 24 years of hands-on retail leadership across E-commerce, Modern Trade, Category Management, and Merchandising Operations, paired with modern technical foundations in Agentic AI, Recommender Systems, and Applied Machine Learning.

My technical architectures are uniquely grounded in a multidisciplinary foundation:

  • Autonomous AI & Global Commerce: Executive Alumnus, Indian Institute of Foreign Trade (IIFT), New Delhi & IIT Madras Pravartak (Professional Certificate in Agentic AI and Data Science).
  • Financial & Economic Engineering: MBA (Finance), Indian Institute of Finance β€” grounding pricing agents in real-world gross margin thresholds, working capital velocity, and P&L governance.
  • Governance & Legal Guardrails: LL.B. (Bachelor of Laws), Faculty of Law, University of Delhi β€” designing hard compliance checks for statutory Maximum Retail Price (MRP) ceilings and contractual vendor agreements.
  • Quantitative First-Principles: B.Sc. (Physics) β€” mathematical modeling, probability distributions, and price elasticity algorithms.
  • Industry Credentials: Certified across Artificial Intelligence, Cloud Infrastructure, and Data Engineering by Google, Microsoft, IBM, and AWS.

Core Value Proposition: Transitioning retail operations from latency-plagued manual reporting into closed-loop autonomous AI systems that enforce margin guardrails, maximize sales velocity, and drive measurable GMV and inventory turnover improvements.

🌐 Unified Autonomous Retail AI Mesh: Closed-Loop Operating Architecture

Rather than isolated point solutions, my flagship neural networks and multi-agent systems interconnect into an enterprise closed-loop quick-commerce intelligence pipeline:

flowchart TD
    classDef vision fill:#EFF6FF,stroke:#3B82F6,stroke-width:2px;
    classDef recsys fill:#F5F3FF,stroke:#8B5CF6,stroke-width:2px;
    classDef pricing fill:#ECFDF5,stroke:#10B981,stroke-width:2px;
    classDef causal fill:#FDF4FF,stroke:#C026D3,stroke-width:2px;
    classDef ops fill:#FFFBEB,stroke:#F59E0B,stroke-width:2px;

    SV["Shelf-Vision<br/><b>(CV Shelf & OOS Audit)</b>"]:::vision
    FG["Fresh-Grade<br/><b>(ViT Quality & Days-to-Spoil)</b>"]:::vision
    UD["Uplift-DRL<br/><b>(DragonNet CATE Targeting)</b>"]:::causal
    BG["BasketGNN<br/><b>(Co-Purchase Graph RecSys)</b>"]:::recsys
    ENS["SASRec & Two-Tower<br/><b>(Sequential & Vector Ensemble)</b>"]:::recsys
    PMO["Perishable-Markdown-Optimizer<br/><b>(Yield & Price Elasticity)</b>"]:::pricing
    OPS["Dark Store Picking Totes<br/><b>& Last-Mile Batching</b>"]:::ops

    SV -- "Real-Time OOS Mask<br/>(Zero-Ghost SKU Filter)" --> BG
    FG -- "FEFO Priority Signal<br/>(SELL_FIRST Boost)" --> BG
    FG -- "Decay Curves & Days<br/>(Markdown Ladder Rules)" --> PMO
    UD -- "Persuadable User IDs<br/>(Targeted Promo Boost)" --> BG
    BG --> ENS
    PMO -- "Dynamic Clearance Prices" --> OPS
    ENS -- "Personalized Replenishment Additions" --> OPS
Loading
  • Zero Ghost-Inventory Recommendations: Shelf-Vision identifies empty out-of-stock shelf facings and broadcasts an OOS mask to BasketGNN, ensuring the recsys never recommends unavailable products to shoppers.
  • FEFO Clearance Acceleration: Fresh-Grade detects produce ripeness and shelf-life, feeding SELL_FIRST priority weights into BasketGNN to stimulate sales velocity while simultaneously triggering dynamic yield markdowns in Perishable-Markdown-Optimizer.
  • Deep Causal Promo Personalization: Uplift-DRL estimates Conditional Average Treatment Effect (CATE) using DragonNet with targeted regularization, filtering out "Sure Things" and "Sleeping Dogs" to direct promotional coupon spend exclusively to "Persuadable" shoppers, feeding high-propensity customer IDs directly into BasketGNN.
  • Multi-Modal RecSys Ensembling: Graph neural networks (BasketGNN) capture high-order co-purchase graph topology, ensembling with causal sequence models (SASRec) and dense vector retrieval (Two-Tower) to drive measurable basket expansion and replenishment conversion.

🧠 Deep Learning Paradigms Mastered

CNN Object Detection ViT Multi-Task Multimodal Fusion Causal Transformer Dual-Encoder Graph Neural Network Temporal Fusion NLP Fine-Tuning Causal CATE

Nine distinct neural architectures shipped to production-grade retail use cases β€” from dense anchor-free object detection on dark-store shelves (YOLOv8), through multi-task Vision Transformer heads for perishable grading (ViT-B/16), graph-propagation collaborative filtering on co-purchase bipartite adjacencies (LightGCN), to deep causal targeted regularization for promo allocation (DragonNet). Each paradigm is matched to the retail problem whose inductive bias it fits best.

πŸš€ Tier-1 Flagship Systems: Multi-Agent & Neural Architectures

System Core Architecture & Stack Business & Technical Impact
Autonomous A2A Vendor Negotiation Arena Streamlit, FastAPI, LangGraph, Groq (openai/gpt-oss-120b), Render Autonomous bilateral procurement bargaining state machine between a Retail Buyer (Milkbasket) and FMCG Supplier with multi-parameter trade-offs (price, volume, SLA) converging within a 6-turn limit. Deployed on Render (FastAPI backend) with an interactive Streamlit arena UI.
Multi-Agent Dynamic Pricing Swarm LangChain, Groq (openai/gpt-oss-120b), Pydantic Autonomous trading desk: Reconciles rival competitor pricing, price elasticity (%Ξ”Q = Ξ΅ Γ— %Ξ”P), statutory MRP ceilings, and margin floors to output executive P&L decision scorecards.
Perishable Markdown & Yield Optimizer LangChain, Groq (openai/gpt-oss-120b), Pydantic Autonomous commercial yield management: Continuously evaluates batch hours-to-expiry against cost salvage thresholds, dynamically triggering aggressive clearance (<24h), promotional markdowns (24-48h), and base price holding (>48h) to eliminate shrinkage.
Autonomous Inventory Optimizer LangGraph, Pydantic, Python Closed-loop replenishment state machine tracking sales velocity, Days of Cover (DoC), vendor MOQs, and lead times to dynamically draft validated JSON Purchase Orders.
VoC Defect Routing Agent LangChain, Groq (openai/gpt-oss-120b), Pydantic Autonomous quick-commerce QA triage: Extracts structured defect assessments from unstructured reviews, classifies failure modes (Temperature, Packaging, Missing Items, Spoilage), rates severity (1-5), and routes tickets to Dark Store Ops, Category Managers, or Supply Chain.
Sequential Recommender (SASRec) PyTorch, Transformers, Causal Self-Attention Captures dynamic user interaction trajectories over time using left-padded causal Transformer self-attention blocks to predict next-basket SKU additions (evaluated on HR@10, NDCG@10).
Retail-SQL-Copilot LangChain, SQLAlchemy, Tool-Calling Autonomous enterprise text-to-SQL copilot with automated schema inspection, query safety guardrails, and natural language executive BI synthesis.
Multimodal SKU Classifier PyTorch, ResNet-18, Hugging Face Hub Late-fusion neural architecture combining vision representations with text encoders (nn.EmbeddingBag) for automated catalog categorization. Model weights published on Hugging Face.
Shelf-Vision: Planogram & OOS Detection Ultralytics YOLOv8, PyTorch, OpenCV, Streamlit Computer vision audit pipeline auditing retail dark-store shelf photos; uses bipartite Hungarian spatial matching to detect empty Out-of-Stock (OOS) slots, misplaced facings, and brand Share-of-Shelf (SoS) %. (Live Interactive App on Streamlit)
Fresh-Grade: ViT Produce Quality & Shelf-Life PyTorch, Vision Transformer (ViT-B/16), timm, Streamlit Multi-task Vision Transformer estimating 4-class perishable produce quality grade alongside continuous days-to-spoil regression (Softplus); chains directly into dynamic markdown price ladders and FEFO dark-store picking. Model weights on Hugging Face. (Live Interactive App on Streamlit)
BasketGNN: Graph Co-Purchase Recommender PyTorch, LightGCN (Sparse Tensors), Streamlit Graph neural network on user-item and SKU co-purchase bipartite graphs; dynamically ingests Shelf-Vision out-of-stock (OOS) masks and Fresh-Grade FEFO perishable priority signals to generate inventory-aware next-basket cross-sell recommendations. (Live Interactive App on Streamlit)
Uplift-DRL: Deep Causal Promo Targeting PyTorch, DragonNet, Targeted Regularization, Streamlit Four CATE estimators (DragonNet, TARNet, T-Learner, S-Learner) with TMLE targeted regularization estimating incremental coupon uplift and sorting shoppers into Persuadables, Sure Things, Lost Causes, and Sleeping Dogs with a budget-aware commercial TargetingBridge. (Live Interactive App on Streamlit)
Last-Mile Delivery Batching Agent LangChain, NetworkX, Groq (gpt-oss-120b) Ingests localized order spikes and utilizes spatial graph logic to autonomously cluster quick-commerce deliveries, adhering to strict rider weight capacities (15kg) and time-to-deliver SLAs.
Enterprise Supply Chain MCP Server FastMCP, Python MCP SDK, SQLite3, JSON-RPC Production Model Context Protocol (MCP) server providing standard JSON-RPC tools over stdio for LLM agents to securely inspect live dark-store inventory levels and vendor SLA lead times.
Agentic REST API Gateway FastAPI, Streamlit, LangChain, Groq (openai/gpt-oss-120b) Enterprise AI API gateway wrapping an autonomous Voice of Customer (VoC) defect triage agent; classifies delivery grievances (Temperature, Packaging, Spoilage), assigns severity 1-5, and routes tickets to Dark Store Ops or Supply Chain. Features an interactive Streamlit console.

πŸ“¦ Extended Portfolio: Specialized Retail AI Microservices

πŸ” Deep Learning, Recommenders & Predictive Forecasting (Click to Expand)
  • Two-Tower Neural Recommendation Engine: Custom dual-encoder neural network in PyTorch with in-batch negative sampling and sub-millisecond FAISS vector indexing for high-throughput candidate retrieval. (Live Interactive App on Streamlit)
  • Demand Sensing & Probabilistic Forecasting (TFT): Temporal Fusion Transformer architecture using PyTorch Forecasting with Quantile Loss ($P_{10}/P_{50}/P_{90}$) and dynamic safety-stock optimization dashboards. (Live Interactive App on Streamlit)
  • Retail Document Intelligence (RAG): Strict enterprise Retrieval-Augmented Generation system using LangChain, ChromaDB, and Gemini 2.5 Flash for grounded citation of standard operating procedures (SOPs) and retail policy documentation. (Live Interactive App on Streamlit)
  • Autonomous Multi-Agent Procurement Squad: Decentralized LangGraph architecture combining real-time web scraping, inventory thresholds, and data-backed vendor negotiation drafting. (Live Interactive App on Streamlit)
βš™οΈ Commercial APIs, Classical ML Benchmarks & Utilities (Click to Expand)

πŸ› οΈ Technical Architecture & Stack

  • Agentic AI & Orchestration: Model Context Protocol (MCP / FastMCP Servers over stdio), LangGraph (Cyclic StateGraphs, Dynamic Routing, Conditional Edges), LangChain, Multi-Agent Swarms, Tool Calling & Function Execution, Structured Output Schemas (Pydantic), Groq (openai/gpt-oss-120b), Google Gemini
  • Deep Learning, Vision & RecSys: Vision Transformers (ViT-B/16 Multi-Task Heads), PyTorch, Ultralytics YOLOv8 (Anchor-Free Object Detection), Bipartite Hungarian Spatial Matching (scipy.optimize.linear_sum_assignment), Self-Attentive Sequential Recommendations (SASRec), Two-Tower Candidate Retrieval, Multimodal Late Fusion (ResNet-18 + Text Encoders), Transformers, Hugging Face Hub
  • Time Series & Predictive Modeling: Temporal Fusion Transformers (TFT), Probabilistic Quantile Loss ($P_{10}/P_{50}/P_{90}$), PyTorch Forecasting, XGBoost, Random Forest, Scikit-Learn
  • Information Retrieval & Vector Search: FAISS (Sub-millisecond Vector Similarity), ChromaDB, Retrieval-Augmented Generation (RAG), Semantic Embedding Search
  • Data Engineering & Analytics: Python, SQL, SQLAlchemy, MySQL, SQLite, Pandas, NumPy, Data Pipeline Optimization, Power BI
  • API & Systems Design: Model Context Protocol (FastMCP / stdio JSON-RPC 2.0), FastAPI, Streamlit, Flask, RESTful Microservices, Automated Workflow Governance & P&L Guardrails

πŸ“« Connect & Collaborate

Popular repositories Loading

  1. Multimodal-SKU-Classifier Multimodal-SKU-Classifier Public

    PyTorch multimodal classification pipeline combining ResNet-18 vision backbone with nn.EmbeddingBag text embeddings to categorize products from images and titles.

    Python 1

  2. Agentic-Inventory-Optimizer Agentic-Inventory-Optimizer Public

    An autonomous supply chain decision agent built with LangGraph, LangChain, and Pydantic that automates retail replenishment and purchase order generation.

    Python 1

  3. Sequential-Recommendation-SASRec Sequential-Recommendation-SASRec Public

    Self-Attentive Sequential Recommendation (SASRec) engine built with PyTorch, causal transformer self-attention blocks, and ranking evaluation (HR@10, NDCG@10)

    Python 1

  4. Shelf-Vision Shelf-Vision Public

    Planogram Compliance & Out-of-Stock (OOS) Detection for retail shelves powered by Ultralytics YOLOv8 and Hungarian spatial bipartite matching.

    Python 1

  5. spendly-expense-tracker spendly-expense-tracker Public

    Full-stack personal finance web application built with Flask to log daily expenses, track budgets, and visualize spending patterns.

    Jupyter Notebook

  6. sbi-financial-database sbi-financial-database Public

    Comprehensive financial database architecture featuring optimized SQL views and an integrated Power BI dashboard for enterprise reporting.

    Jupyter Notebook