neuralwithmohit
  • About
  • Blog
  • AI System Design
  • Cloud Comparison
Subscribe
neuralwithmohit

Content

  • Blog
  • AI System Design
  • Case Studies
  • Cloud Comparison

Site

  • About
  • Pricing
  • Resume

Elsewhere

  • LinkedIn
  • YouTube
  • HuggingFace

© 2026 Mohit Kumar Dubey. Built with Next.js & FastAPI.

On this page

  • About
  • Selected work
  • Skills
  • Background

About

Mohit Kumar Dubey

Head of AI, ML & Engineering · Entermind Malaysia · Kuala Lumpur, Malaysia

I'm an AI and engineering leader with 16+ years building AI-first products and the organizations that ship them. I'm currently Head of AI, ML & Engineering at Entermind, a Kuala Lumpur consultancy serving enterprise and government clients across SEA and MENA, where I own the company-wide AI and engineering roadmap and lead an 11-engineer team across backend, AI/ML, frontend and QA.

Before Entermind I spent nearly a decade at VerSe Innovation (Dailyhunt / Josh / BuddyLoan), India's largest short-video and news platform at 350M+ users. I grew from Lead Software Engineer to Director of Engineering, leading a 35+ engineer organization across AI/ML, backend, mobile and web, and owning the AI/ML roadmap across content and lending. Earlier, I delivered enterprise and government mobility across the GCC and platform applications for SmartTV, Xbox and banking-grade security.

My work sits at the intersection of production GenAI delivery and engineering leadership: I own engineering P&L, design the operating model, manage executive stakeholders across the Gulf, SEA and the US, and still architect the systems myself. The projects below are written in STARL form (Situation, Task, Action, Result, Learning) so the reasoning, not just the result, is visible.

Download résuméGet in touch

Selected work

Projects, in STARL

Situation, Task, Action, Result, Learning, so the reasoning is visible, not just the outcome.

Entermind

Mindara, AI Employee for Enterprise Engagement

Visit
Situation
Enterprise engagement started with static intake forms that captured shallow requirements and lost nuance before sales ever engaged.
Task
Replace the form with an always-on AI colleague that surfaces the real requirement and hands sales a structured brief.
Action
Architected across seven capability levels (Level 1 in production, voice staged) on a LiteLLM gateway with Claude Opus / OpenAI / Gemini, LangGraph orchestration, Langfuse observability and ElevenLabs voice.
Result
15% increase in solutioning lead calls.
Learning
The model's value is in the follow-up, not the first answer. Adaptive questioning elicits intent that a form never will.
  • LiteLLM
  • Claude Opus
  • Gemini
  • Langfuse
  • ElevenLabs
  • LangGraph
Entermind

Rukun Ready, Guardrail LLM for Policy Validation

Visit
Situation
Enterprise clients needed LLM outputs checked against national policy principles before deployment, with no off-the-shelf guardrail aligned to Malaysia's Rukun Negara.
Task
Build a guardrail model that scores outputs for compliance and returns structured verdicts with safe rewrites.
Action
Fine-tuned Qwen2.5-32B with LoRA/PEFT to emit structured JSON compliance verdicts, served on vLLM/RunPod, and released it open-source on HuggingFace.
Result
Firm's first open-source AI asset; 3K+ downloads; a standing pre-sales credibility anchor.
Learning
A small, sharply-scoped fine-tune shipped openly can build more credibility than a larger closed system.
  • Qwen2.5-32B
  • LoRA
  • PEFT
  • vLLM
  • RunPod
  • HuggingFace
Entermind

AI Digital Twin, First Abu Dhabi Bank

Client · confidential
Situation
The Chief Credit Officer's calendar was the bottleneck for routine credit queries.
Task
Give stakeholders 24/7 access to that expertise without adding to the CCO's meeting load.
Action
Built a multimodal conversational twin over enterprise knowledge, with voice, memory and agentic workflows, on Azure AI Foundry, LangGraph, Tavus/ElevenLabs/Pipecat, pgvector, Redis and Azure AI Search.
Result
20% lower routine meeting load and 40% better information availability.
Learning
Digital twins win when they absorb the routine load, freeing the human expert for the genuine exceptions.
  • Azure AI Foundry
  • LangGraph
  • Tavus
  • Pipecat
  • pgvector
  • Azure AI Search
Entermind

GEO Platform, GXS (Grab Group)

Client · confidential
Situation
Buyers had shifted discovery to AI assistants where the brand was largely invisible.
Task
Make the brand measurably more visible and citable inside AI-driven discovery.
Action
Built citation-aware analytics and optimization across four assistants (ChatGPT, Gemini, Claude, Perplexity) with competitor benchmarking on AWS Bedrock, model APIs, LangGraph and PostgreSQL.
Result
45% lift in AI brand citation scores.
Learning
"SEO for AI answers" is a real, measurable surface. You can't optimize what you don't first instrument.
  • AWS Bedrock
  • OpenAI
  • Gemini
  • Claude
  • Perplexity
  • LangGraph
VerSe Innovation

BuddySmart, Conversational AI & Agentic RAG

Visit
Situation
Loan discovery was form-driven with heavy drop-off.
Task
Turn discovery into a conversation that matches users to loans and actually converts.
Action
Built a conversational matchmaking assistant with real-time eligibility checks and personalized offers on LangGraph/LangChain, Cohere embeddings, Qdrant, AWS Bedrock and ECS.
Result
25% disbursal-conversion uplift and 20% retention gain.
Learning
Conversational eligibility removes the biggest drop-off point: users abandon forms, not conversations.
  • LangGraph
  • LangChain
  • Cohere
  • Qdrant
  • AWS Bedrock
  • ECS
VerSe Innovation

DailyTatva RAG v1 & v2

Visit
Situation
Personalized news retrieval at 350M+ users, with LLM spend growing faster than usage.
Task
Hold retrieval quality high while bending the cost curve.
Action
Built the v1 retrieval/inference foundation, then added model routing and response caching in v2 on LangGraph/LangChain, Gemini, Cohere embeddings, Vertex AI, Qdrant and ECS.
Result
30–40% cut in LLM API cost while improving retrieval quality.
Learning
At scale, routing and caching move cost more than model choice. Most requests don't need the biggest model.
  • LangGraph
  • Gemini
  • Cohere
  • Vertex AI
  • Qdrant
  • ECS

Skills

Technical skills

AI & ML
LLMs, RAG, Agentic RAG, Multi-Agent Systems, Fine-tuning (LoRA, PEFT), NLP, Recommendation Systems, Evaluation Harness, Agentic Loop design, Prompt Engineering, MLOps, Credit-Risk Models, Structured Outputs
Agent Frameworks
LangChain, LangGraph, CrewAI, DeepAgents
Training & Serving
PyTorch, TensorFlow, vLLM, SGLang, FastAPI
Model Platforms
AWS Bedrock, Amazon SageMaker, Vertex AI, Azure AI Foundry, OpenAI, Gemini, Claude, Cohere
Voice & Multimodal
ElevenLabs, Tavus, LiveKit, Pipecat, Daily, Serper, Tavily
Data & Vector Stores
PostgreSQL, pgvector, Qdrant, MongoDB, Chroma, FAISS, Redis
MLOps & Observability
Langfuse, Phoenix, Opik, LiteLLM, MLflow, Kubeflow, Airflow
Infrastructure
AWS, GCP, Docker, Kubernetes, ECS, EC2, CI/CD
Languages
Python, Pandas, NumPy, Pydantic, SQL, Swift

Background

Education & certifications

Education

  • Executive Post Graduate Programme in Machine Learning & AI (Specialization: Generative AI)
    International Institute of Information Technology (IIIT), Bangalore
    Nov 2024
  • B.Sc., Information Technology
    Kanpur University, India
    Mar 2009

Certifications

  • AWS Certified AI Practitioner
  • Responsible AI and Advanced Generative AI (DeepLearning.AI)
  • Mastering System Design
  • Software Architecture: Patterns for Developers
  • SOLID Principles in Software Design
  • REST-based Microservices API Development in Golang
Awards

Integrated Team Award · Team Excellence (×2) · Reward & Recognition (VerSe)

Languages

English, Hindi

Exposure

Malaysia, Singapore, UAE, India

← Back home