Sanyam is available for hire
Hire SanyamSanyam Jain
Verified Expert in Engineering
Data Scientist and Developer
Noida, Uttar Pradesh, India
Toptal member since July 27, 2026
Bio
Sanyam builds GenAI systems that survive production: LLM agents, RAG pipelines, and real-time voice assistants on live enterprise data. He architected a governed conversational BI platform in which the LLM never holds database credentials and shipped two production voice agents that handle live customer calls in seven languages. With 6 years across GenAI and applied ML, plus 50+ independent client engagements at a 100% success rate, Sanyam owns problems end-to-end.
Portfolio
JK Tyre
OpenAI, Tesseract OCR, Databricks, Amazon S3 (AWS S3)...
An Online Freelance Agency
Python 3, FastAPI, LangGraph, LangChain, Agentic RAG Systems, LLM Agents...
Cashify
Scikit-learn, Neo4j, Random Forests, Gradient Boosting, XGBoost, Clustering...
Experience
- Python - 7 years
- FastAPI - 5 years
- Artificial Intelligence (AI) - 5 years
- LangChain - 3 years
- AI Voice Agents - 3 years
- Model Context Protocol (MCP) - 3 years
- LangGraph - 3 years
- Agentic AI - 2 years
Work Experience
Lead Data Scientist
2024 - PRESENT
JK Tyre
- Architected and deployed a production platform enabling business teams to query the data warehouse and internal documents in natural language, enforcing data security independently of the LLM with a fail-closed SQL firewall and RBAC layer.
- Shipped 2 production voice AI agents on LiveKit answering live dealer calls in seven Indian languages with mid-call language switching and nine integrated API tools, freeing roughly 40% of sales-officer call-handling bandwidth.
- Delivered a daily warranty-claims forecasting system across around 6,000 dealers using an exposure-adjusted LightGBM Tweedie count model, cutting operational false positives 80% (760 to 152 per day) and lifting precision from 41.5% to 58.8%.
- Extended the BI platform with document Q&A via RAG over PDFs, combining native text extraction, vision-model parsing, and OCR fallback with permission-scoped retrieval, enterprise SSO through OIDC, and full query audit logging.
- Engineered the shared real-time voice framework underneath both agents: barge-in handling, voice-activity detection, telephony noise cancellation, SIP human handover, call recording to S3, and LLM transcript summarisation.
- Built a 5-model recommendation engine on PySpark, processing 14.5 million transactions per run, segmenting 5,900 dealers with DBSCAN auto-tuned across 3,200 configurations to drive cross-sell for 3,600+ dealers and 1,700 extra orders daily.
- Diagnosed a 3x over-prediction traced to rolling-window non-stationarity and corrected it to 0.99x accuracy using a 48-month history window, seasonal retraining, and out-of-sample calibration, validated by rolling-origin cross-validation.
- Built dealer incentive-scheme simulators with dealer-response modeling on a SAP to S3 to Databricks ETL pipeline, served as FastAPI microservices integrated into the enterprise ERP, delivering around 17% sales uplift in pilot schemes.
- Automated monthly route planning for field sales officers using K-Means and DBSCAN clustering with greedy TSP routing over a Haversine travel-time matrix under 8-hour-day constraints, cutting travel time and improving coverage.
Technologies: OpenAI, Tesseract OCR, Databricks, Amazon S3 (AWS S3), Model Context Protocol (MCP), Microsoft Entra ID, LightGBM, Unity Catalog, LiveKit, FAISS, MongoDB, PySpark, Scikit-learn, DBSCAN, Delta Lake, SciPy, FastAPI, LangGraph, LangChain, Retrieval-augmented Generation (RAG), LLM Agents, LLM Fine-tuning, LLM Reasoning, LLM Integration, Large Language Models (LLMs), Agentic AI, SQL, PostgreSQL, Deepgram, Azure Text to Speech, Cartesia, Speech-to-Text (STT), Text-to-Speech (TTS), Conversational AI, Docker, Git, REST APIs, WebSockets, Pydantic, Microservices, OAuth 2, OpenID Connect (OIDC), Pandas, NumPy, Spark, Delta Live Tables (DLT), ETL, Data Pipelines, Amazon EC2, AWS Lambda, XGBoost, Clustering, K-means Clustering, Time Series Forecasting, CI/CD Pipelines, Feature Engineering, Model Deployment, Recommendation Systems, Prompt Engineering, Vector Databases, Embedding Models, Natural Language Processing (NLP), AI Agents, Python, Artificial Intelligence (AI), API Integration, Product Development, AI Integration, RAG Systems, Agentic RAG Systems, RAG Pipelines, Agentic AI Systems, Agentic Frameworks, Agentic Workflow Design, Agentic Coding, ChatGPT, Anthropic, Claude, Amazon Web Services (AWS), Software Architecture, AI Automation, SaaS, AI Product Management, Databases, Integrations, Workflow, Workflow Automation, NoSQL, Architecture, Full-stack Development, Cloud Native, Distributed Systems, Full-stack, Data Validation, Data Science, GitHub Actions, APIs, Machine Learning, Machine Learning Operations (MLOps), AI Architecture, Technical Leadership, Data Engineering, Automation, OpenAI API, Data Modeling, Data Processing, Data Mining, Security, Back-end, Back-end Development, Software Development, AI Agent Orchestration, System Architecture, Authentication, Generative Artificial Intelligence (GenAI), Debugging, Team Leadership, Azure, Azure Blob Storage, Claude API, Microsoft 365, OAuth, Role-based Access Control (RBAC), Best Practices, Database Migration, Pipelines, Serverless, Optical Character Recognition (OCR), API Development, Amazon Bedrock
Freelance Data Scientist & AI Engineer
2021 - 2025
An Online Freelance Agency
- Delivered 50+ freelance projects for international clients with a 100% Job Success Score, spanning AI/ML engineering, LLM integrations, data science, and automation across multiple industries.
- Built custom GenAI solutions for clients, including RAG pipelines, LLM-powered applications, and AI agents, taking each engagement from problem scoping through delivery as the sole engineer.
- Developed machine learning and data-analysis solutions for business problems, translating ambiguous client requirements into working, production-ready deliverables on tight timelines.
- Maintained long-term client relationships through reliable delivery and clear communication, earning repeat engagements and consistent top ratings across projects.
- Designed and shipped REST APIs and back-end services with Python and FastAPI to serve ML and LLM features, handling data pipelines, validation, and integration with clients' existing systems.
- Acted as sole technical owner on most engagements, making architecture, tooling, and deployment decisions independently while communicating trade-offs clearly to non-technical stakeholders and founders.
- Earned consistent repeat business and top ratings through reliable delivery and responsive communication, maintaining long-term relationships with clients across multiple countries.
- Built automation and workflow solutions that eliminated manual client processes, combining API integrations, scheduled jobs, and AI agents to streamline operations and reduce operational overhead.
Technologies: Python 3, FastAPI, LangGraph, LangChain, Agentic RAG Systems, LLM Agents, OpenAI, AI Agents, Prompt Engineering, PostgreSQL, Machine Learning, Natural Language Processing (NLP), Databases, Data Science, Databricks, REST APIs, Model Context Protocol (MCP), Python, Deep Learning, PyTorch, Recommendation Systems, Transformer Models, Llama 3, Data Mining, Agentic AI, Data Modeling, Data Processing, Azure, Azure Blob Storage, Claude API, Microsoft 365, OAuth, Best Practices, API Development
Data Scientist
2021 - 2023
Cashify
- Built a dynamic pricing engine for refurbished consumer electronics using price-elasticity and demand modeling, automatically refreshing prices weekly across 360 SKUs and improving gross margin by roughly 12% across the catalog.
- Delivered a customer-similarity and fraud-detection service from scratch in two months, combining a Neo4j graph of around 100,000 nodes with a Random Forest classifier to link related accounts far beyond simple phone-number matching.
- Optimized fraud-detection inference to 130 milliseconds average latency, enabling real-time screening inside the live transaction flow rather than as an offline batch review.
- Engineered the feature pipelines and automated the end-to-end pricing workflow in Python, removing manual price-setting across the catalog and making weekly repricing fully hands-off.
- Modeled customer relationships as a graph to cluster connected accounts, surfacing coordinated fraud rings that identifier-based matching could not detect on its own.
- Partnered with product and engineering to integrate both ML services directly into live sales operations, owning scoping, modeling, deployment, and monitoring as the sole data scientist on each workstream.
Technologies: Scikit-learn, Neo4j, Random Forests, Gradient Boosting, XGBoost, Clustering, Feature Engineering, Model Evaluation, Model Deployment, Supervised Learning, Anomaly Detection, Graph Databases, Cypher, Network Analysis, Fraud Detection, Dynamic Pricing, Price Analysis, SQL, Pandas, NumPy, ETL, Data Pipelines, REST APIs, Low Latency, Git, AWS Lambda, Python, Artificial Intelligence (AI), API Integration, Product Development, AI Integration, ChatGPT, Amazon Web Services (AWS), AI Automation, Databases, Integrations, NoSQL, Architecture, Cloud Native, Distributed Systems, Data Validation, Data Science, GitHub Actions, APIs, Machine Learning, Machine Learning Operations (MLOps), AI Architecture, Automation, Data Modeling, Data Processing, Data Mining, Microsoft 365, OAuth, Best Practices, Serverless, API Development
Data Scientist
2020 - 2021
MapMyIndia
- Built XORAT, an intelligent search-autocomplete system for geospatial queries, applying BERT-based named entity recognition and spaCy dependency parsing to interpret user location intent and improve suggestion relevance.
- Developed Woodpecker, an NLP module parsing raw unstructured addresses into structured components such as city, state, road, and point-of-interest, using custom NER and chunk parsing to raise geocoder precision.
- Trained and evaluated custom named-entity recognition models on Indian address and place-name data, handling transliteration, abbreviations, and inconsistent formatting common in real-world user input.
- Collaborated with the mapping and search engineering teams to move NLP models from prototype into the live geocoding stack serving consumer and enterprise map products.
Technologies: BERT, SpaCy, NER, Natural Language Processing (NLP), Named-entity Recognition (NER), Tokenization, Text Classification, Information Extraction, Transformers, Hugging Face Transformers, Hugging Face, Natural Language Toolkit (NLTK), PyTorch, Scikit-learn, Geospatial Data, GIS, Geocoding, Pandas, NumPy, Python, Artificial Intelligence (AI), API Integration, AI Integration, Amazon Web Services (AWS), Databases, Integrations, NoSQL, Data Validation, Data Science, APIs, Machine Learning, Machine Learning Operations (MLOps), Automation, Data Modeling, Data Processing, Data Mining, OAuth, Best Practices, Optical Character Recognition (OCR), API Development
Project History
Governed Conversational BI Platform
Business teams waited days for analysts to answer routine data questions, and the obvious fix, letting an LLM query the warehouse directly, was blocked by governance: row-level rules meant an APAC rep must never read EMEA data, however the question was phrased. My job was to make that a yes.
I built a platform where the LLM generates SQL but never executes it and never holds credentials. Every query is parsed with sqlglot into an AST and rejected fail-closed on any write, DDL, or session-mutating node, including nodes buried in CTEs, subqueries, or UNIONs. Surviving queries pass through a multi-pass RBAC rewriter, injecting per-user row filters and column masks with server-bound values, and remain correct across JOINs and UNIONs. The agent runs in a per-user subprocess whose only data egress is a token-scoped MCP tool. Document Q&A runs on a parallel RAG plane with hybrid PDF extraction, cosine-similarity vector search, and permission-scoped retrieval, secured by OIDC with PKCE via Microsoft Entra ID.
BUSINESS IMPACT
Multi-day analyst request cycles became self-serve answers in minutes for hundreds of staff, releasing the analytics team for higher-value work while satisfying enterprise data-governance and audit requirements.
I built a platform where the LLM generates SQL but never executes it and never holds credentials. Every query is parsed with sqlglot into an AST and rejected fail-closed on any write, DDL, or session-mutating node, including nodes buried in CTEs, subqueries, or UNIONs. Surviving queries pass through a multi-pass RBAC rewriter, injecting per-user row filters and column masks with server-bound values, and remain correct across JOINs and UNIONs. The agent runs in a per-user subprocess whose only data egress is a token-scoped MCP tool. Document Q&A runs on a parallel RAG plane with hybrid PDF extraction, cosine-similarity vector search, and permission-scoped retrieval, secured by OIDC with PKCE via Microsoft Entra ID.
BUSINESS IMPACT
Multi-day analyst request cycles became self-serve answers in minutes for hundreds of staff, releasing the analytics team for higher-value work while satisfying enterprise data-governance and audit requirements.
Real-time Multilingual Voice AI Agents
Sales officers were losing hours a day to routine inbound calls, breakdown calls went unanswered outside business hours, leaving customers stranded, and the customer base spoke seven different languages, which no single support desk could cover.
I built a shared real-time voice framework and two production agents on it. Myra takes roadside breakdown calls in Hindi, Marathi, Bengali, Gujarati, Tamil, Telugu, and Kannada with live mid-call STT/TTS language switching on DTMF keypress, running a deterministic twelve-state flow from vehicle verification through GPS capture over WhatsApp, nearest pitstop dispatch, and ticket creation, with locks preventing duplicate tickets. Navya authenticates dealers via OTP, then answers live account questions via 9 API tools covering credit, payments, orders, statements, and stock, plus RAG over product documentation. The framework handles barge-in, voice-activity detection, telephony noise cancellation, call recording to S3, LLM summarization, MongoDB logging, and SIP human handover.
BUSINESS IMPACT
Roughly 40% of sales-officer bandwidth was freed for selling, round-the-clock breakdown coverage, and 7-language reach without hiring a multilingual call center.
I built a shared real-time voice framework and two production agents on it. Myra takes roadside breakdown calls in Hindi, Marathi, Bengali, Gujarati, Tamil, Telugu, and Kannada with live mid-call STT/TTS language switching on DTMF keypress, running a deterministic twelve-state flow from vehicle verification through GPS capture over WhatsApp, nearest pitstop dispatch, and ticket creation, with locks preventing duplicate tickets. Navya authenticates dealers via OTP, then answers live account questions via 9 API tools covering credit, payments, orders, statements, and stock, plus RAG over product documentation. The framework handles barge-in, voice-activity detection, telephony noise cancellation, call recording to S3, LLM summarization, MongoDB logging, and SIP human handover.
BUSINESS IMPACT
Roughly 40% of sales-officer bandwidth was freed for selling, round-the-clock breakdown coverage, and 7-language reach without hiring a multilingual call center.
Dealer Claims Forecasting at Scale
The claims team was receiving around 760 alerts a day, most of which were wrong. Alert fatigue meant the genuine warranty risk was being ignored, and planners had no reliable forward view of claim volume across roughly 6,000 dealers of varying sizes.
I formulated the problem as an exposure time hazard. An empirical survival and hazard curve, convolved with sales history, produces an at-risk population, which is fed as a log-exposure offset into a LightGBM Tweedie count model so that predictions scale correctly with dealer size, alongside a separate classifier for claim acceptance. Validation used leakage-safe rolling-origin cross-validation behind a Poisson-deviance gate against three baselines. I diagnosed a roughly 3x over-prediction traced to rolling-window non-stationarity and corrected it to 0.99x accuracy using a 48-month window, seasonal retraining, and out-of-sample calibration. Shipped as a scheduled, credential-free Databricks job feeding Tableau.
Business impact: false positives were cut roughly 80%, from 760 to 152 a day, and precision rose from 41.5% to 58.8%, so the team could act on alerts again and plan warranty provisioning against a trusted three-day forecast.
I formulated the problem as an exposure time hazard. An empirical survival and hazard curve, convolved with sales history, produces an at-risk population, which is fed as a log-exposure offset into a LightGBM Tweedie count model so that predictions scale correctly with dealer size, alongside a separate classifier for claim acceptance. Validation used leakage-safe rolling-origin cross-validation behind a Poisson-deviance gate against three baselines. I diagnosed a roughly 3x over-prediction traced to rolling-window non-stationarity and corrected it to 0.99x accuracy using a 48-month window, seasonal retraining, and out-of-sample calibration. Shipped as a scheduled, credential-free Databricks job feeding Tableau.
Business impact: false positives were cut roughly 80%, from 760 to 152 a day, and precision rose from 41.5% to 58.8%, so the team could act on alerts again and plan warranty provisioning against a trusted three-day forecast.
Multi-tenant AI Support Assistant API
Support agents were drafting replies from memory against long policy documents, so quality varied by agent and onboarding a new client company meant engineering work every time.
I built a FastAPI service where each company owns its own MCP server exposing its guidelines as a resource plus its own action tools. A LangGraph agent runs a triage, respond, tools, finalize flow: it classifies the message into a guideline category, pulls targeted per-category guidance through a tool call rather than stuffing the whole playbook into context, optionally invokes company tools within a bounded loop, then drafts the recommended reply. Every node and tool call appends to a persisted trace, so any run is inspectable afterwards alongside latency and category metrics. The provider sits behind a factory keeping the agent model-agnostic, and tests run fully offline against a fake model.
Business impact: agents get on-policy draft replies in seconds instead of hunting through documentation, quality stops depending on who picks up the ticket, and onboarding a new company is a config change rather than a development cycle.
I built a FastAPI service where each company owns its own MCP server exposing its guidelines as a resource plus its own action tools. A LangGraph agent runs a triage, respond, tools, finalize flow: it classifies the message into a guideline category, pulls targeted per-category guidance through a tool call rather than stuffing the whole playbook into context, optionally invokes company tools within a bounded loop, then drafts the recommended reply. Every node and tool call appends to a persisted trace, so any run is inspectable afterwards alongside latency and category metrics. The provider sits behind a factory keeping the agent model-agnostic, and tests run fully offline against a fake model.
Business impact: agents get on-policy draft replies in seconds instead of hunting through documentation, quality stops depending on who picks up the ticket, and onboarding a new company is a config change rather than a development cycle.
Education
2016 - 2020
Bachelor's Degree in Mechanical Engineering
Abes Engineering College - Ghaziabad, Uttar Pradesh, India
Skills
Libraries/APIs
PySpark, Scikit-learn, SciPy, SpaCy, XGBoost, REST APIs, Pydantic, Pandas, NumPy, Hugging Face Transformers, Natural Language Toolkit (NLTK), PyTorch, OpenAI API, API Development, Claude API
Tools
Git, Claude Code, Named-entity Recognition (NER), ChatGPT, Claude, GIS, Geocoding, Tableau
Languages
SQL, Python 3, Python, Cypher
Frameworks
LangGraph, OAuth 2, Spark, Delta Live Tables (DLT), Agentic Frameworks, LightGBM, Apache Spark
Paradigms
Model Context Protocol (MCP), ETL, Microservices, Automation, Best Practices, Anomaly Detection, Test-driven Development (TDD), Role-based Access Control (RBAC)
Platforms
Databricks, LiveKit, Docker, Amazon EC2, AWS Lambda, Amazon Web Services (AWS), Cloud Native, Azure
Storage
Amazon S3 (AWS S3), MongoDB, PostgreSQL, Data Pipelines, Databases, NoSQL, Data Validation, Neo4j, Database Migration, Microsoft Entra ID, Graph Databases
Other
OpenAI, FAISS, DBSCAN, Delta Lake, FastAPI, BERT, NER, Retrieval-augmented Generation (RAG), LangChain, Deepgram, Cartesia, Clustering, Machine Learning, Deep Learning, LLM Agents, LLM Fine-tuning, LLM Reasoning, LLM Integration, Large Language Models (LLMs), Agentic AI, Azure Text to Speech, Speech-to-Text (STT), Text-to-Speech (TTS), Conversational AI, K-means Clustering, Feature Engineering, Model Deployment, Recommendation Systems, Prompt Engineering, Vector Databases, Embedding Models, Natural Language Processing (NLP), Random Forests, Gradient Boosting, Model Evaluation, Supervised Learning, Fraud Detection, Low Latency, Tokenization, Text Classification, Information Extraction, Transformers, Hugging Face, Application Security, AI Voice Agents, AI Agents, Artificial Intelligence (AI), API Integration, Product Development, AI Integration, Agentic AI Systems, Agentic Workflow Design, Agentic Coding, Anthropic, AI Product Management, Integrations, Workflow, Workflow Automation, Data Science, APIs, Data Mining, Data Modeling, Data Processing, Generative Artificial Intelligence (GenAI), AI-assisted Development, AI Agent Orchestration, AI Engineering, Back-end, Back-end Development, Software Development, Debugging, Code Review, Streaming, Software Engineering, Microsoft 365, OAuth, Pipelines, Serverless, Unity Catalog, CI/CD Pipelines, Statistics, WebSockets, Dynamic Pricing, Price Analysis, Geospatial Data, Voice Activity Detection (VAD), Session Initiation Protocol (SIP), RAG Systems, Agentic RAG Systems, RAG Pipelines, Software Architecture, AI Automation, SaaS, Architecture, Full-stack Development, Distributed Systems, Full-stack, GitHub Actions, Machine Learning Operations (MLOps), AI Architecture, Technical Leadership, Data Engineering, Event-driven Systems, Cloud Architecture, Observability, Security, System Architecture, Authentication, Team Leadership, Azure Blob Storage, Optical Character Recognition (OCR), Amazon Bedrock, Tesseract OCR, OpenID Connect (OIDC), Time Series Forecasting, Network Analysis, Transformer Models, Llama 3, Actuarial, Cursor AI
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