Kiran Dapkar
Verified Expert in Engineering
Data Scientist and Developer
Pune, Maharashtra, India
Toptal member since January 27, 2025
Kiran Dapkar is a senior AI engineer who builds production-grade LLM platforms and agent systems with a strong focus on governance, safety, and scale. He has hands-on experience designing high-throughput LLM pipelines, policy-driven tool execution, and controlled ChatGPT traffic routing for enterprise environments. Known for turning complex AI ideas into reliable systems, Kiran helps teams deploy AI that is both powerful and responsible.
Portfolio
Experience
- Python 3 - 4 years
- Data Science - 4 years
- Machine Learning - 4 years
- Prompt Engineering - 3 years
- Retrieval-augmented Generation (RAG) - 3 years
- AI Agents - 1 year
- LangGraph - 1 year
- Agentic AI - 1 year
Preferred Environment
MacOS, Windows, Outlook, Slack, Excel 365, Python 3, Bash, Pandas, Microsoft PowerPoint, Large Language Models (LLMs)
The most amazing...
...solution I've built was a governed LLM platform with LangGraph agents and Data Shapley, reducing data size by 44% and saving $1.1 million.
Work Experience
AI Engineer
DAXA
- Developed LLM usage analytics APIs aggregating 10,000+ calls/day, powering dashboards for model usage, token consumption, violations, and top users; reduced dashboard latency by 60% via optimized SQL aggregations.
- Designed and implemented an MCP server integrated with LangGraph, developing a chatbot application that leverages multi-agent orchestration and tool routing for Atlassian services.
- Developed LangGraph-based agents (Ask-Pebblo) with caching, memory for chat history, React-style tool routing for LLM-based governance and support flows.
- Architected and owned a high-throughput LLM proxy and analytics platform, processing 50,000+ LLM calls/day with <50ms latency overhead, including streaming response capture and token accounting for billing and compliance.
- Designed and owned a Model Context Protocol (MCP)-based policy enforcement engine enabling ordered, pluggable policy evaluation (READ/WRITE/DELETE permissions, PII detection, server allowlisting) with <10ms enforcement overhead.
- Built real-time LLM policy violation tracking pipelines logging to PostgreSQL and Elasticsearch, enabling enterprise monitoring for command injection, unsafe tool usage, and sensitive data leaks.
- Improved tool classification accuracy and performance by iterating LLM prompts (confidence scoring, reasoning fields) and introducing a Redis cache-first architecture, reducing DB queries by 80% and enforcement latency from 50ms to 5ms.
- Led end-to-end migration from Proxima Applications to Pebblo Agents, refactoring core data, API, and analytics layers, and resolving production inconsistencies from MongoDB to PostgreSQL transitions.
Senior Data Scientist and AI/ML Engineer
American Express
- Developed a generative-AI-powered insight engine using LLMs, Pandas, and ReAct prompting to analyze high-balance defaulter spend patterns, streamlining risk model analysis.
- Streamlined case study tools at American Express, significantly improving operational efficiency.
- Built a generative AI chatbot using DPR, LLMs, and prompt engineering, boosting response rates by 80% and achieving a 92% success rate.
- Benchmarked LLM reasoning on tabular data, achieving a 15% improvement over GPT 3.5 by leveraging zero-shot, few-shot, and chain-of-thought prompting.
- Reduced risk model population by 44% through Data Shapley, cutting costs by 50% and achieving a $1.1 million financial impact while improving model performance.
- Published award-winning research on GBDT-Shapley at the Odyssey AI-ML Summit, earning the Global Decision Science SVP Award for innovation and collaboration at American Express.
- Enhanced HPT standards by improving model selection stability, optimizing log loss, and introducing structural complexity to mitigate data drift in production systems.
Data Scientist and AI/ML Engineer
HDFC Bank
- Developed a robust optical character recognition (OCR) pipeline using EasyOCR for text extraction from electricity bills, achieving structured analysis despite high bill heterogeneity.
- Implemented an automated solution for transaction data extraction from banking texts, incorporating advanced feature engineering techniques.
- Enhanced the model accuracy benchmark for transaction extraction by 22% through fine-tuning with BERT and advanced NLP techniques.
- Achieved >95% coverage and 97%+ accuracy in classifying banking transaction categories (derogatory vs non-derogatory, cash deposits, credit card payments) via regex, text mining, and frequency analysis.
- Built a multilingual abusive voice detection system leveraging acoustic features (MFCC, pitch, STFT) and Conv1D neural networks, enabling automated flagging of abusive calls in collections.
- Optimized large-scale address matching pipelines by refactoring algorithms with RAPIDS cuDF (GPU-accelerated), reducing processing time from 60+ hrs to 25 hrs on 10M+ records.
- Automated cropping pattern identification using Sentinel-2 NDVI indices, FarmBeats API, and regex-driven pipelines for agricultural insights.
- Evaluated employee performance in retail banking using Decision Tree models, analyzing PB Reconnect program scorecards to identify training effectiveness and key performance drivers.
Experience
AVAA - AI Voice Appointment Assistant
https://aiassistantspeech.netlify.app/UHC Policy Chatbot
Speech-to-Indian Sign Language Conversion
https://github.com/kirandapkar/Speech-to-Indian-Sign-Language-Conversion/Education
Bachelor's Degree in Computer Science
Indian Institute of Technology - Mumbai, India
Certifications
Microsoft Certified: Azure Data Fundamentals
Microsoft
Neural Networks and Deep Learning
Coursera
AI for Everyone
Coursera
Computer Vision: Object Detection with OpenCV and Python
Coursera
Analyze Text Data with Yellowbrick
Coursera
Skills
Libraries/APIs
Pandas, XGBoost, Hugging Face Transformers, Natural Language Toolkit (NLTK), OpenAI API, OpenCV, TensorFlow, PyTorch, OpenAI Assistants API, React
Tools
Slack, Microsoft PowerPoint, Claude
Languages
Python 3, Bash, Bash Script, C++, Regex, Python
Frameworks
Django, LangGraph, Flask
Platforms
MacOS, Windows, Google Cloud Platform (GCP), Azure, Docker, Amazon Web Services (AWS), LiveKit, Netlify
Paradigms
Model Context Protocol (MCP)
Storage
Redis
Other
Outlook, Large Language Models (LLMs), Data Science, Machine Learning, Natural Language Processing (NLP), Prompt Engineering, BERT, Text Mining, Artificial Intelligence (AI), Agentic AI, Excel 365, Transformers, Retrieval-augmented Generation (RAG), Mixtral, Chatbots, Data Shapley, EasyOCR, Deep Neural Networks (DNNs), Deep Learning, LangChain, FastAPI, Open-source LLMs, AI Agents, Cursor AI, APIs, Computer Science, Optical Character Recognition (OCR), Speech-to-Text (STT), Text-to-Speech (TTS), OpenAI, RAG Pipelines, ChromaDB, Embedding Models, Deepgram, OpenRouter, Supabase, Railway
How to Work with Toptal
Toptal matches you directly with global industry experts from our network in hours—not weeks or months.
Share your needs
Choose your talent
Start your risk-free talent trial
Top talent is in high demand.
Start hiring