Bhoumik is available for hire
Hire BhoumikBhoumik Shah
Verified Expert in Engineering
Operations Research Scientist and Developer
Bengaluru, Karnataka, India
Toptal member since September 13, 2022
Bio
Bhoumik is a machine learning and operations research specialist with six years in AI, notably in ads moderation, linehaul cost optimization, and McKinsey's routing tool. He's excelled in cost reduction and system efficiency, adept with jsprit, OR-Tools, and ML model development. With both a bachelor's and master's degree in mechanical engineering, with a focus on operations research, Bhoumik also holds certifications in deep learning and related fields.
Portfolio
CogniSphere
Agentic AI, Agentic Frameworks, Agentic RAG Systems, AI Agents, AI Chatbots...
Ryan Mc Andrew
Artificial Intelligence (AI), Large Language Models (LLMs), OpenAI, Python...
Car Guy AI, LLC
Large Language Models (LLMs), Python, OpenAI, Natural Language Processing (NLP)...
Experience
- Operations Research - 6 years
- Supply Chain Optimization - 6 years
- Vehicle Routing - 6 years
- Vehicle Routing Problem (VRP) - 6 years
- Jsprit - 5 years
- Machine Learning - 4 years
- Google OR-Tools - 3 years
- Data Science - 2 years
Work Experience
Founder
2026 - PRESENT
CogniSphere
- Founded CogniSphere (https://cognisphere.sh), an AI engineering practice that builds custom agents and software for businesses on an open-source agent platform I designed.
- Architected and open-sourced CogniSphere Harness (MIT, TypeScript/Node), a multi-agent orchestration server: https://github.com/cognisphere-sh/cognisphere-harness.
- Designed durable execution with per-agent SQLite queues, threads that run serially within a thread and concurrently across threads, mid-run steering, and evidence-based resume or resend after failures.
- Built a single plugin interface for Telegram, Gmail/Google Workspace, cron scheduling, agent-to-agent messaging, and HTML artifact publishing, extensible to any CRM or API.
- Enabled agents to turn plain-language feedback into persistent knowledge, instructions, and versioned skills, so business users change behavior without developers.
- Deployed a 4-agent team for a Texas plant broker: Daisy (supplier intelligence), Rosie (sales and orders), Lily (requirements), and Nova (platform engineering).
- Built a supplier network of 1,395 supplier locations and a 48K+ row plant catalogue, ingested from supplier emails, PDFs, spreadsheets, images, and websites.
- Ran the agents in production: in one audited week, Daisy handled 158 threads and ~80 supplier emails, and Rosie processed 173 operator messages and 63 supplier portal events.
- Designed a closed improvement loop: Lily turns feedback into requirements, Nova ships and deploys approved app changes, and business agents update their own procedures.
Technologies: Agentic AI, Agentic Frameworks, Agentic RAG Systems, AI Agents, AI Chatbots, AI Programming, AI Prompts, TypeScript, Node.js, SQLite, OpenAI, AI Security, RAG Systems, GraphRAG, Access Control, Data Architecture, AI Governance, LangGraph, Amazon Bedrock, Model Context Protocol (MCP), AI Evaluation, Agentic AI Systems, REST APIs, Claude Agent SDK
AI Developer with AI agent development experience for construction recruitment firm
2026 - 2026
Ryan Mc Andrew
- Replaced a manual JobAdder workflow, where staff hand-keyed emailed resumes and jobs, with a custom ATS and autonomous AI recruiter for a Canadian construction recruitment firm of 40+ recruiters.
- Built the ATS on Supabase (PostgreSQL, Auth) with a responsive, mobile-friendly Next.js frontend and Google OAuth sign-in, improving on JobAdder's UI/UX.
- Built Dory, an AI recruiting agent, on my open-source CogniSphere harness (https://github.com/cognisphere-sh/cognisphere-harness), with her own Gmail inbox and Telegram account.
- Automated intake: Dory monitors recruiter emails, extracts resumes and job requirements from attachments and message bodies, and writes structured records to the ATS.
- Designed a 10-step matching pipeline that combines keyword search, semantic search, and RAG to shortlist candidates for each new job and jobs for each new candidate.
- Added LLM-based deep evaluation: Dory reads candidates' past roles and projects, and researches client companies and LinkedIn profiles online, to judge fit beyond keyword overlap.
- Enabled recruiters to email Dory follow-up requests for deeper evaluations, refined searches, and record updates, so she works as a teammate rather than a search tool.
- Gave admins real-time Telegram access to Dory for chat, admin tasks, and recurring reports scheduled through the harness's cron scheduler.
Technologies: Artificial Intelligence (AI), Large Language Models (LLMs), OpenAI, Python, Machine Learning, Natural Language Processing (NLP), Supabase, Vercel, Next.js, AI Agents, Agentic AI, Agentic Frameworks, Agentic RAG Systems, AI Chatbots, AI Security, AI Programming, LangGraph, Amazon Bedrock, Model Context Protocol (MCP), AI Evaluation, Agentic AI Systems, REST APIs, Claude Agent SDK
[Toptal Teams] AI Engineer with Chatbot experience
2026 - 2026
Car Guy AI, LLC
- Built an autonomous AI sales agent that engages dealership leads over SMS, email, and web chat, answers inventory and vehicle questions, and drives in-person dealership visits.
- Rebuilt a year-long, team-built LangGraph lead-conversion agent from scratch on my open-source CogniSphere harness (https://cognisphere.sh), shipping to production in 2 weeks.
- Enabled dealership staff to change agent behavior in real time through plain-language feedback that updates its SOPs, so behavior changes no longer need a developer.
- Replaced rigid, scripted graph workflows with an open-ended agent that handles off-script questions and multi-step tasks where the previous system failed.
- Hardened security by sandboxing the agent and screening every message from untrusted channels for prompt injection, quarantining malicious inputs before they reach the agent.
- Deployed the production stack on AWS (EC2, Lambda, RDS), with SES for email and Twilio for SMS, delivering a live multi-channel system within the 2-week timeline.
- Integrated the agent into the dealership's website as a live chat interface, letting shoppers talk to it directly alongside the text and email channels.
- Try it yourself: go to https://www.billcoleautomallbluefield.com/ --> click on chat button on bottom left --> Give your phone number and email Id --> The agent will send you a text message. This is deployed in production.
Technologies: Large Language Models (LLMs), Python, OpenAI, Natural Language Processing (NLP), Chatbots, Chatbot Conversation Design, Technical Leadership, Model Context Protocol (MCP), AI Evaluation, Agentic AI Systems, REST APIs, Claude Agent SDK
ML & Analytics Expert | Food Industry
2024 - 2024
New Taiwan Trading Corp DBA NTFOODS
- Developed end-to-end recommendation engine for a platform that sells raw materials to restaurants.
- Scrapped restaurants' menus and used items in the menu to generate personalized recommendations for raw materials.
- Deployed recommendation engine as a service on AWS using AWS Lambda, DynamoDB, and Step Functions.
Technologies: Machine Learning, Python, Data Modeling, Chatbots, ChatGPT, Food, Grocery Delivery, eCommerce, Analytics, OpenAI, Natural Language Understanding (NLU), API Integration, Node.js, Claude, ChatGPT API, AI Data Classification, AI Modeling, Data Processing, AutoGen, CrewAI, Beautiful Soup, Speech Recognition, Agentic AI, n8n, Agentic Frameworks, PostgreSQL, Architecture, CTO, Fractional CTO, Security, Supabase, Supabase Auth, Claude Code, Databases, JavaScript, HubSpot, HubSpot CRM, Agentic RAG Systems, Claude API, OpenAI API, User Interface (UI), AI Evaluation, Agentic AI Systems, REST APIs
AI Consultant
2024 - 2024
Mercatus Center at George Mason University - Main
- Built a heuristic algorithm to identify and extract relevant information and amendments from policy documents published by US federal agencies. Due to large size of documents (300+ pages), it could not be done directly LLMs.
- Used LLMs to make amendments into federal regulations based on published policy by multiple federal agencies.
- Built a front-end application using Streamlit to compare changes in regulation.
Technologies: Machine Learning, Web Scraping, Java Natural Language Processing (JNLP), Artificial General Intelligence (AGI), Deep Learning, SQL, Natural Language Processing (NLP), Artificial Intelligence (AI), Large Language Models (LLMs), AI Prompts, Large Data Sets, Finance, Capital Markets, Entity Extraction, AI Agents, OpenAI, Natural Language Understanding (NLU), API Integration, Google Cloud, Claude, ChatGPT API, CrewAI, Beautiful Soup, Agentic AI, n8n, Agentic Frameworks, PostgreSQL, Architecture, CTO, Fractional CTO, Security, Supabase, Supabase Auth, Claude Code, Integrations, JavaScript, HubSpot CRM, Agentic RAG Systems, Claude API, OpenAI API, WhatsApp API, User Interface (UI), REST APIs
AI/ML Engineer
2024 - 2024
Gold Coin Group LLC
- Developed an end-to-end application for scrapping scanned images of handwritten mails from an internal website and extracting data from it.
- Used AWS textract, and OpenAI along with Donut model to do OCR of handwritten images and extract information in structured format.
- Implemented web scraping solution using Selenium and Beautiful Soup.
Technologies: Computer Vision, Python, OpenCV, TensorFlow, Pandas, Machine Learning, Artificial Intelligence (AI), Deep Learning, Workflow Automation & System Integration, Natural Language Processing (NLP), Optical Character Recognition (OCR), Keras, AI Prompts, Large Data Sets, Finance, Capital Markets, OpenAI, Natural Language Understanding (NLU), API Integration, Google Cloud, Claude, Tesseract, ChatGPT API, AI Data Classification, AI Modeling, Data Processing, AutoGen, Beautiful Soup, Agentic AI, Replit, Agentic Frameworks, PostgreSQL, Security, JavaScript, Claude API, User Interface (UI), REST APIs
Applied Scientist
2021 - 2023
Amazon India
- Developed automated moderation systems to streamline the review of more than 5 billion ads annually.
- Led a team of three in enhancing current machine learning systems, cutting down the number of ads requiring manual moderation by 60%.
- Achieved a $500 million annual reduction in shipping invoice estimation errors by training a gradient boosting machine (GBM) model using H2O to handle a large training dataset of 250 million samples.
Technologies: Artificial Intelligence (AI), H20, Machine Learning, Gradient Boosting, Spark, PyTorch, Generative Pre-trained Transformers (GPT), Natural Language Processing (NLP), Data Science, Deep Learning, Large Language Models (LLMs), Amazon Web Services (AWS), Data Analysis, Predictive Modeling, OpenAI GPT-3 API, Technical Leadership, Haystack, Retrieval-augmented Generation (RAG), Trend Analysis, Advertising Technology (Adtech), Machine Learning Operations (MLOps), Neural Networks, Algorithms, Retail & Wholesale, Inventory, CSV File Processing, Full-stack, Web Development, Amazon Bedrock AgentCore, Generative Pre-trained Transformer 3 (GPT-3), Consulting, Leadership, SQL, Product Management, PySpark, Data Science Product Manager, GitLab, AI Programming, AI Model Training, Pandas, NumPy, NoSQL, FastAPI, Neo4j, Graph Databases, Python 3, Models, Generative Artificial Intelligence (GenAI), Scikit-learn, Classifier Development, Supervised Learning, Teamwork, Regression, Data Engineering, Web Scraping, Spark ML, Optical Character Recognition (OCR), Workflow Automation & System Integration, TensorFlow, eCommerce, Analytics, Sentiment Analysis, Modeling, Data Collection, Big Data, C#, Document Parsing, Minimum Viable Product (MVP), Automation, Microsoft Excel, AI Prompts, Large Data Sets, Java, Entity Extraction, AI Agents, Automatic Speech Recognition (ASR), Natural Language Understanding (NLU), API Integration, Node.js, AI Data Classification, AI Modeling, Data Processing, Speech Recognition, Agentic AI, Agentic Frameworks, PostgreSQL, Data Pipelines, Automotive, WhatsApp API, REST APIs, Excel 365
Operations Research Scientist
2019 - 2021
Amazon India
- Built a dispatch time optimizer for the Amazon India network to deliver 10% of shipments one day faster. Used concepts of multi-agent learning and mixed-integer linear programming to solve the NP-complete problem.
- Used concepts of vehicle routing optimization and built a line-haul route planner tool for the Amazon line-haul network. It combined multiple lanes to increase vehicle utilization. This resulted in an annual freight cost reduction of $10 million.
- Created a workforce scheduling tool using constraint programming to automate and optimize human resources deployment at the Amazon sort center, which resulted in an 8% reduction in the workforce.
Technologies: Operations Research, Optimization, Vehicle Routing, Network Optimization, Python, Java, Gurobi, CPLEX, Amazon Web Services (AWS), Data Analysis, Data Science, Time Series, Predictive Modeling, Technical Leadership, Machine Learning, Trend Analysis, Machine Learning Operations (MLOps), Neural Networks, Algorithms, Retail & Wholesale, Inventory, CSV File Processing, Consulting, SQL, Product Management, PySpark, Data Science Product Manager, GitLab, AI Model Training, Pandas, NumPy, NoSQL, FastAPI, Python 3, Models, Location Services and Maps, Spatial Databases, Classifier Development, Teamwork, Regression, Data Engineering, Web Scraping, Software Architecture, Workflow Automation & System Integration, eCommerce, Generative Pre-trained Transformers (GPT), Analytics, Grocery Delivery, Modeling, Data Collection, Big Data, Document Parsing, Minimum Viable Product (MVP), Automation, Microsoft Excel, API Integration, Data Processing, PostgreSQL, Data Pipelines, Automotive, REST APIs, Excel 365
Knowledge Analyst
2017 - 2019
McKinsey & Company
- Developed a state-of-the-art tool to solve vehicle routing problems with real-life constraints using heuristic algorithms. Deployed the solution as a web application and implemented it in over 15 client scenarios spanning various industries.
- Collaborated with multiple global clients for topology and production planning optimization.
- Developed a routing algorithm for solving ship container routing for a leading petrochemical producer.
Technologies: Supply Chain, Supply Chain Optimization, Supply Chain Management (SCM), Data Analysis, Data Science, Time Series, Forecasting, Predictive Modeling, Technical Leadership, Machine Learning, Trend Analysis, Finance, Neural Networks, Algorithms, Retail & Wholesale, Inventory, CSV File Processing, Consulting, SQL, Product Management, Data Science Product Manager, GitLab, AI Programming, AI Model Training, Pandas, NumPy, NoSQL, FastAPI, Python 3, Models, Location Services and Maps, Spatial Databases, Classifier Development, Supervised Learning, Teamwork, Regression, Data Engineering, Web Scraping, Spark ML, Software Architecture, Workflow Automation & System Integration, Analytics, Grocery Delivery, Modeling, Data Collection, Big Data, Automation, Microsoft Excel, Large Data Sets, Java, AI Data Classification, AI Modeling, Data Processing, Automotive Systems, REST APIs, Excel 365
Project History
Automatic Grading of Handwritten Tests
The automatic grading system is an innovative tool designed to assess handwritten answer sheets with high precision. Utilizing state-of-the-art optical character recognition (OCR) technology, the system deciphers handwritten text and efficiently converts it into digital format. Beyond this conversion, smart post-processing ensures accuracy, addressing common issues associated with handwriting variations. Once the answers are digitized, the system employs large language models (LLMs) to evaluate the content, ensuring a comprehensive understanding of the answer's context and relevance. The integration of these technologies provides a swift and accurate grading process, reducing manual effort and increasing consistency in evaluation.
Routing Optimization
I built a state-of-the-art routing optimization tool using jsprit, OR-Tools, and a proprietary algorithm to optimize outbound routing for an eCommerce giant. The tool resulted in an annual cost reduction of USD 10 million.
Ads Moderation System
Developed automated moderation systems to streamline the review of over 5 billion ads per year. Also, I led a team of three in enhancing current machine learning systems, cutting down the volume of ads requiring manual moderation by 60%.
Virtual Companion
The Telegram virtual companion bot is a groundbreaking chatbot leveraging the power of large language models (LLMs) to foster human-like interactions. With the integration of a long-term memory feature, the bot not only engages in real-time conversations but also personalizes interactions over time based on previous exchanges. Hosted on AWS Lambda for efficient serverless deployment, this bot ensures seamless and scalable interactions. Its exceptional conversational capabilities have resonated with users, amassing over 5,000 active users within just the first week of its launch.
YT Chat
I built a chatbot that enables users to chat with any YouTube channel. I built both the front and the back end.
Stack:
• Front end: React, Tailwind, and shadcn
• Authentication: Clerk
• Database: MongoDB
• Back end: Python with FastAPI
• RAG: LlamaIndex, Pinecone vector database, and OpenAI
Stack:
• Front end: React, Tailwind, and shadcn
• Authentication: Clerk
• Database: MongoDB
• Back end: Python with FastAPI
• RAG: LlamaIndex, Pinecone vector database, and OpenAI
Real-time Transcription for Medical Industry
A full-stack, real-time transcription service. The product records conversations between doctors and patients in real time.
Stack:
• Front end: React, Tailwind, and shadcn
• Back end: Python and FastAPI
• Transcription: AssemblyAI
• Summarisation: OpenAI
Stack:
• Front end: React, Tailwind, and shadcn
• Back end: Python and FastAPI
• Transcription: AssemblyAI
• Summarisation: OpenAI
Education
2012 - 2017
Master's Degree in Mechanical Engineering
Indian Institute of Technology Bombay - Mumbai, India
Certifications
AUGUST 2018 - PRESENT
Deep Learning
Coursera
JULY 2018 - PRESENT
Structuring Machine Learning Projects
Coursera
MAY 2018 - PRESENT
Introduction to Data Science in Python
Coursera
Skills
Libraries/APIs
Pandas, NumPy, Scikit-learn, OpenCV, Beautiful Soup, Claude API, OpenAI API, REST APIs, React, Spark ML, TensorFlow, Google Vision API, WhatsApp API, PyTorch, Telegram Bot API, PySpark, Java Natural Language Processing (JNLP), Keras, Node.js
Tools
Google OR-Tools, ChatGPT, GitLab, Microsoft Excel, AI Prompts, Claude, Claude Code, Claude Agent SDK, PyCharm, IntelliJ IDEA, Haystack, Azure OpenAI Service, n8n, Gurobi, CPLEX, Amazon SageMaker, Shadcn, GraphRAG
Languages
Python, SQL, Python 3, HTML, Java, TypeScript, JavaScript, C#
Frameworks
Agentic Frameworks, Next.js, Flask, LlamaIndex, Selenium, AutoGen, LangGraph, Spark, Tailwind CSS, Scrapy
Paradigms
Spatial Databases, Automation, Model Context Protocol (MCP)
Platforms
CrewAI, Amazon Web Services (AWS), Docker, Azure Functions, Google Cloud Platform (GCP), Replit, H20, AWS Lambda, Azure, HubSpot, Vercel
Industry Expertise
Retail & Wholesale, Automotive
Storage
PostgreSQL, Databases, Amazon DynamoDB, NoSQL, Neo4j, Graph Databases, Google Cloud, MongoDB, Data Pipelines, SQLite
Other
Jsprit, Operations Research, Supply Chain, Supply Chain Optimization, Supply Chain Management (SCM), Machine Learning, Optimization, Vehicle Routing Problem (VRP), Vehicle Routing, Data Science, Natural Language Processing (NLP), Artificial Intelligence (AI), OpenAI, Large Language Models (LLMs), Data Analysis, Prompt Engineering, Technical Leadership, Retrieval-augmented Generation (RAG), Neural Networks, Algorithms, Inventory, CSV File Processing, Amazon Bedrock AgentCore, Generative Pre-trained Transformer 3 (GPT-3), Consulting, Leadership, Data Science Product Manager, AI Programming, AI Model Training, FastAPI, Models, Location Services and Maps, Generative Artificial Intelligence (GenAI), LoRa, Hugging Face, Classifier Development, Supervised Learning, Teamwork, Regression, Pinecone, Data Engineering, Artificial General Intelligence (AGI), Data Scraping, Chatbots, Open-source LLMs, Workflow Automation & System Integration, eCommerce, Analytics, APIs, Vectorization, Sentiment Analysis, Modeling, Data Collection, Big Data, Minimum Viable Product (MVP), Large Data Sets, Entity Extraction, Natural Language Understanding (NLU), API Integration, ChatGPT API, AI Data Classification, AI Modeling, Data Processing, Agentic AI, Architecture, Fractional CTO, Supabase, Supabase Auth, Workflow Automation, Integrations, Workflow, Troubleshooting, Agentic RAG Systems, AI Evaluation, Agentic AI Systems, Excel 365, Simulations, Machine Learning Operations (MLOps), LangChain, Time Series, Predictive Modeling, OpenAI GPT-3 API, OpenAI GPT-4 API, Full-stack Development, AI Chatbots, SaaS, Trend Analysis, Finance, Advertising Technology (Adtech), Full-stack, Product Management, Chatbot Conversation Design, Web Scraping, PDF Scraping, Software Architecture, Embeddings from Language Models (ELMo), Fine-tuning, Computer Vision, Large Language Model Operations (LLMOps), Grocery Delivery, Document Parsing, Gemini, Capital Markets, AI Agents, Automatic Speech Recognition (ASR), Tesseract, Speech Recognition, Cursor AI, CTO, Automotive Systems, User Interface (UI), AI Governance, Amazon Bedrock, Deep Learning, Convolutional Neural Networks (CNNs), Recurrent Neural Networks (RNNs), Network Optimization, Generative Pre-trained Transformers (GPT), Gradient Boosting, Optical Character Recognition (OCR), Forecasting, Web Development, Legal Technology (Legaltech), Travel, Speech-to-Text (STT), Google Speech-to-Text, Data Modeling, Food, Security, HubSpot CRM, AI Security, RAG Systems, Access Control, Data Architecture
Collaboration That Works
How to Work with Toptal
Toptal matches you directly with global industry experts from our network in hours—not weeks or months.
1
Share your needs
Discuss your requirements and refine your scope in a call with a Toptal domain expert.
2
Choose your talent
Get a short list of expertly matched talent within 24 hours to review, interview, and choose from.
3
Start your risk-free talent trial
Work with your chosen talent on a trial basis for up to two weeks. Pay only if you decide to hire them.
Top talent is in high demand.
Start hiring