Summary
Overview
Work History
Education
Skills
Tool Use
Websites
Timeline
Generic

M.F Ammar

Dehiwala

Summary

Lead Data Scientist and NLP/LLM architect with deep experience building retrieval-augmented, agentic AI systems and production-grade workflows. Expert in multi-agent orchestration (CrewAI, Agno), end-to-end LLM pipeline automation (n8n, Langflow), and observability/tracing of assistant behavior (LangSmith, Langfuse). Proven track record in designing, fine-tuning, and deploying context-aware AI applications using HuggingFace model hub, LlamaIndex, LangChain, and LLMs (OpenAI GPT, Anthropic Claude, OpenRouter models), delivering measurable accuracy and latency improvements.

Overview

4
4
years of professional experience

Work History

Lead Data Scientist

Veracity Group
Colombo
06.2023 - Current
  • Designed and deployed workflow automation pipelines for data ingestion, LLM orchestration, and downstream inference using n8n and Langflow, enabling visual, composable agent-flows and reducing manual orchestration overhead.
  • Architected and led multi-agent AI systems, leveraging CrewAI and AgnoAGI to decompose complex tasks into collaborating, specialized agents, improving throughput and autonomous decision-making.
  • Built and integrated observability and traceability layers across assistant applications with LangSmith and Langfuse, capturing full interaction traces (inputs, tool calls, prompt versions, latencies, and outcomes) to accelerate failure triage and iterative prompt/system refinement.
  • Developed high-fidelity RAG applications using LlamaIndex, AgnoAGI, and LangChain to ground LLM responses in external knowledge sources, increasing contextual relevance and reducing hallucination.
  • Standardized and executed model fine-tuning pipelines (including PEFT techniques like QLoRA and quantization), tracked experiments with Weights & Biases (W&B), and leveraged the Hugging Face Model Hub for selecting and adapting foundation models.
  • Delivered measurable improvements: e.g., double-digit percentage gains in prediction accuracy, and significant latency reductions through combined architectural and prompt/system refinements.
  • Provide technical guidance and mentorship to the NLP team, assisting with problem-solving, algorithm/model selection, system design, and training/fine-tuning.
  • Collaborate with cross-functional teams to identify opportunities for NLP and GenAI applications, and provide expertise on the feasibility and implementation of NLP solutions.

Data Scientist

VeracityAI
Colombo
06.2021 - 06.2023
  • Application of various machine learning algorithms, like text analytics, natural language processing (NLP), regression models, neural networks, and clustering, using PyTorch.
  • Developed and performed transfer learning techniques for Hugging Face Transformer and Longformer-based pre-trained models.
  • Performed hyperparameter tuning and built my own model training and optimization pipelines using the PyTorch and Weights & Biases platforms to fine-tune and optimize 'Text to Text Transfer' Transformers (T5, T5-v1) and BERT-based models.
  • Extensively worked with the Hugging Face Model Hub and fine-tuned various pre-trained models for text generation, text augmentation, and question answering downstream tasks.
  • Led the upgrade of an existing project's model to the RoBERTa model, significantly improving system performance.
  • Achieved a 96% increase in model accuracy after finetuning, and an 80% reduction in response time, enhancing user experience and improving learning outcomes.
  • Spearheaded the optimization and fine-tuning of GPT-3 models (Ada, Davinci) to enhance prediction accuracy in agricultural yield forecasting.
  • Implemented innovative prompt engineering techniques, resulting in a 98% improvement in prediction accuracy with an external data source. Contributed to the development of optimized agricultural practices through accurate crop yield predictions.
  • Implemented GPT-3.5 and GPT-4 models with the integration of LangChain and LLMA Index using OpenAI API to enhance the power of external data sources (RAG).
  • Tested and implemented various open-source LLMs, such as Mistral, Phi 3, LLaMa3, etc. Using Ollama and Groq, with text-to-speech and speech-to-text models, to enhance the user experience.

Trainee Data Scientist

VeracityAI
Colombo
03.2021 - 06.2021
  • Perform data cleaning, feature scaling, and feature engineering using Pandas and NumPy in Python.
  • Create and design reports that will use gathered metrics to infer and draw logical conclusions from past and future behavior.
  • Work with the NLTK library for NLP data processing and finding the patterns.
  • Implemented AllenNLP for a downstream question-answering task with BIDAF and ELMO word embeddings separately.
  • Responsible for developing system models, prediction algorithms, solutions to prescriptive analytics problems, and data mining techniques.

Education

MSc - Artificial Intelligence

University of Moratuwa
08-2026

BSc (Hons) - Software Engineering

University of the West of England
Bristol, UK

Skills

  • Artificial intelligence (AI)
  • Natural language processing (NLP)
  • Machine learning (ML)
  • Deep Learning (DL)
  • Reinforcement Learning (RL)(PPO, DPO, GRPO etc)
  • Fine-tuning Pipelines and Optimization
  • Prompt Engineering and Prompt Tuning
  • Parameter Efficient Fine Tuning (PEFT)(LoRA, QLoRA etc)
  • Text Generation, Text Augmentation, Question Answering
  • Developing LLM-based Agentic RAG Applications
  • Technical leadership and mentorship
  • Collaboration with Cross-functional Teams
  • AI Research and Feasibility Studies

Tool Use

  • Retrieval-augmented generation (RAG) orchestration using LlamaIndex, AgnoAGI, and LangChain
  • Chainlit, Streamlit, and Gradio for prototyping and quick development
  • Utilizing UV for modern development
  • Multi-agent system design and orchestration with CrewAI and AgnoAGI
  • Model integration - OpenAI, Anthropic, Groq, Mistral, OpenRouter, LiteLLM, Hugging Face, Ollama models
  • Visual and programmatic pipeline construction using n8n, Flowise, and Langflow for modular LLM-driven flows
  • Interaction tracing, prompt/version tracking, and debugging using LangSmith and Langfuse (hierarchical run views, metrics, prompt/version diffing, and reranking analysis)
  • MLOps utilizing Weights and Biases

Timeline

Lead Data Scientist

Veracity Group
06.2023 - Current

Data Scientist

VeracityAI
06.2021 - 06.2023

Trainee Data Scientist

VeracityAI
03.2021 - 06.2021

MSc - Artificial Intelligence

University of Moratuwa

BSc (Hons) - Software Engineering

University of the West of England

Similar Profiles

CREATE PROFILE
M.F Ammar