Touhidul Alam

GenAI Solution Architect | NLP Specialist

Munich, Germany, CET

Touhi

About

As a Senior AI Solution Architect, I design and deliver enterprise-grade Generative AI systems, from scalable RAG reference architectures to production LLM evaluation frameworks. At Accenture, I lead technical delivery for on-prem/sovereign AI initiatives, drive client workshops and hackathons to validate GenAI use cases, and mentor engineers on architecture decisions balancing security, compliance, and scale.

Work Experience

Accenture
Munich
Frankfurt

2020 - Present

GenAI Solution Architect

Tech Lead for enterprise on-prem/sovereign AI delivery — owned solution architecture and platform engineering, defining a scalable RAG reference architecture and operational standards for rollout. Built an evaluation & optimization workflow for LLM applications, introducing an eval framework and observability to track quality, latency, and cost in production. Shipped an NLP recommendation system to production, improving content relevance and user engagement. Core member of Tech Innovation, running client discovery workshops, leading internal hackathons, and driving rapid prototyping (PoCs/MVPs) to validate AI use cases. Mentored junior engineers and guided key architecture decisions, keeping delivery aligned with security, compliance, and scalability constraints.

Fraunhofer IIS
Erlangen

2019 - 2020

Research Assistant

Developing and integrating a Deep learning based temporal tagger for Voice Assistants as a part of Master thesis.

Accenture
Munich

2018 - 2019

Student Intern

Developed a 3D holographic chatbot, increasing user engagement through interactive experiences.

Smart Digital GmbH
Gerlingen

2017 - 2018

Student Assistant

Implemented a real-time anomaly detection system using time-series data to monitor and alert on operational discrepancies, enhancing system reliability

Panacea Live Limited
Dhaka

2016 - 2017

Backend Developer

Developed the Panacea Web system. Explored potential business use cases for fraud detection technology.

Education

University of Stuttgart

2017 - 2020
Master's Degree in Computational Linguistics

Islamic University of Technology

2011 - 2015
Bachelor's Degree in Computer Science and Engineering

Skills

Python
NLP
ML
LLM
RAG
Fine-tuning
Multi-agent Systems
LLM Evaluation
Langchain
Haystack
Langfuse
Vector DB
Graph DB
FastAPI
SQL
Cypher
Pytorch
Azure
AWS
Docker
Kubernetes
CI/CD
MLOps

Projects

A Voice-based LLM Crew Agent

A text/voice-based multi-agent system with RAG in specialized domain

Python
OpenAI
Langchain
CrewAI
Streamlit
Elevenlabs
Crawler

LLM Multi-Agent for Search

Simple multi-agent system with external search API to search, analyze, generate content

Python
OpenAI
Langchain
Langgraph
Streamlit

Knowledge Graph LLM Agent

A knowledge-graph based LLM agent to perform search with different Cypher query

Python
OpenAI
Neo4J
Langchain
Streamlit

Tags generation

Simple prototyping of generating tags from forums with Open-source LLM

Python
LLM
Mistral-7B

DA-Time

A domain-adapted temporal expression recognizer for the English Voice Assistant domain

Python
Bert
Keras

AI Music Generation

A simple language modeling project to generate music trained by MIDI files pattern

Python
Language Modeling
Pytorch

Dialogue Act Classification

A Simple Dialogue Act Classification project using Keras from speech and text with an Ensemble model

Python
Keras
Ensemble model

Emotion Classification from Twitter text and image

Document Level Emotion Classification from Twitter Text and Images

Python
Perceptron

Minimalistic Language Modeling

A Simple n-gram language modeling from scratch to understand the basic concept of language modeling

Python

Publications

Enhancing Pipeline-Based Conversational Agents with Large Language Models

This study examines enhancing pipeline-based conversational agents with GPT-4, focusing on a hybrid approach that combines LLM strengths with existing system safeguards.

Python
GPT-4

New Domain, Major Effort? How Much Data is Necessary to Adapt a Temporal Tagger to the Voice Assistant Domain

The study introduces DA-Time, a hybrid temporal tagger for voice assistants, proving that minimal in-domain data via transfer learning significantly enhances temporal expression recognition in diverse domains.

Python
BERT
Keras

PÂTÉ: A Corpus of Temporal Expressions for the In-car Voice Assistant Domain

This study introduces a crowdsourcing method to collect and annotate natural-language commands with temporal expressions for AI voice assistants, differing from typical domains, to enhance their interaction with applications.

Python