Koushik Deb is an AI engineer, AI researcher and author. He builds Generative-AI and Agentic-AI systems at Accenture, and his doctoral research at IIIT Kalyani studies how large language models behave across English, Hindi and Bengali. His wider work covers machine learning, natural language processing and multilingual AI.
Koushik Deb is an AI Engineer and Tech Lead at Accenture, where he designs and ships Generative-AI and Agentic-AI systems for global healthcare clients. In parallel he is a PhD candidate at IIIT Kalyani, researching bias detection and mitigation in multilingual large language models across English, Hindi and Bengali.
Across thirteen years at Accenture, Cognizant, IBM and Capgemini he has built software for Apple, IKEA, Cigna, Roche, Tiffany & Co. and Sanofi, and spent three and a half years as an Assistant Professor teaching NLP and machine learning. He holds three granted patents, has published twelve papers, and in September 2026 published his first book, Keep It Simple: AI for Teachers.
📍 Kolkata, West Bengal, India
Thirteen years of building software, most of it with machine learning or language technology at the centre. The tools have changed from classic models to generative systems, but the habit has not: decide what a system should do, then measure whether it does.
At Accenture, Koushik designs and delivers Generative-AI and Agentic-AI solutions for healthcare clients. The work covers retrieval-augmented generation, agent workflows and the engineering that turns a proof of concept into a product people use.
RAG explained in plain language →NLP runs through most of his career: ML/NLP roles at IBM and Cognizant, three and a half years teaching it at IEM Kolkata, and published work on sentiment, social-media threat detection and mental-health text. One granted patent covers several NLP tasks running offline on one device.
Offline NLP patent →His doctoral work studies how commercial LLMs respond to biased framing in prompts. The flagship study tested six commercial models over 683,616 responses; confirmation-pattern prompts cut neutral answers by 19.42 percentage points on average.
Read the study explainer →Before LLMs came applied machine learning: fraud detection, disease prediction from drug information, and RKDOS, a density-based outlier detection algorithm that won a best paper award at SKIMA 2018.
RKDOS explainer →Most model testing happens in English. His PhD research at IIIT Kalyani checks whether a model's behaviour holds in Hindi and Bengali as well, and the flagship study found vulnerability to biased prompts was higher in Hindi and Bengali than in English.
Why the language of the prompt matters →In both roles the recurring question is whether an AI system can be trusted for a particular job. In research that means controlled, repeatable measurement; at work it means testing a model on real tasks before it reaches users.
How to evaluate an LLM at work →Koushik started in 2013 as a software engineer at Capgemini, building .NET applications for clients such as Sanofi Aventis. The move into machine learning came at IBM in 2017, where he built ML and NLP solutions for Tiffany & Co. In parallel he completed an Executive M.Tech at IIT (ISM) Dhanbad in 2018.
That year he joined the Institute of Engineering & Management, Kolkata, as an Assistant Professor. Three and a half years of teaching and research in NLP, sentiment analysis and topic modelling produced most of his early publications, and both of his 2019 patent applications were filed during this period.
He returned to industry in 2022 with ML/NLP work for Apple and IKEA at Cognizant, then moved to Generative AI at Accenture. The same year he began a PhD at IIIT Kalyani. The two tracks feed each other. Client work shows where language models fail in practice, and the research measures those failures carefully, especially outside English.
Teaching left one more habit: explaining technical ideas plainly. That habit became his first book, Keep It Simple: AI for Teachers, and the free AI for Teachers guides on this site.
AI Engineer · Tech Lead, GenAI
Senior Developer — ML/NLP
Assistant Professor · Computer Science
Developer — ML/NLP
Software Engineer
Koushik's research background is in NLP and applied machine learning, and his doctoral work asks whether a language model's fairness survives translation. The flagship paper measured cognitive-bias vulnerability in six commercial LLMs across three languages over 683,616 responses. Earlier work spans social-media threat detection, mental-health NLP, opinion mining and density-based outlier detection.
IIIT Kalyani
Bias detection and mitigation in multilingual large language models. Supervisor: Dr. Imon Mukherjee; co-supervisor: Dr. Debarshi Kumar Sanyal (IACS, Kolkata). Registration PHD/R/21003/651.
IIT (ISM) Dhanbad
CGPA 9.03.
Govt. College of Engineering & Textile Technology, Serampore (WBUT)
CGPA 8.43.
Email koushikdeb88@gmail.com for research collaboration, GenAI work or speaking invitations.