NLP Engineer

What a NLP Engineer does daily

• LLM fine-tuning and adaptation — adapting pre-trained large language models to specific domains and tasks using supervised fine-tuning (SFT) and preference optimisation (RLHF, DPO — Direct Preference Optimisation); parameter-efficient fine-tuning with LoRA (Low-Rank Adaptation) and QLoRA (quantised LoRA) to fine-tune models on consumer hardware; fine-tuning open-source models (LLaMA 3, Mistral, Gemma) for Sinhala language understanding, legal document analysis, medical report summarisation, or customer service routing; the core production NLP engineering task in 2026
• RAG (Retrieval-Augmented Generation) system development — the most widely deployed LLM application architecture; combining a vector database (Pinecone, Weaviate, Chroma, Qdrant) with an LLM to answer questions from a company's private document corpus; document ingestion, chunking strategies, embedding model selection (OpenAI text-embedding, Sentence Transformers, BGE models); retrieval pipeline evaluation; reranking (ColBERT, cross-encoder reranking); LangChain, LlamaIndex, or Haystack as orchestration frameworks; deployed as internal knowledge base Q&A, customer support, compliance document search
• Prompt engineering and LLM evaluation — designing system prompts and few-shot examples that reliably elicit high-quality outputs from LLMs; prompt chaining and agent architectures (LangChain Agents, AutoGen, CrewAI); structured output generation (Instructor library, Pydantic validation of LLM outputs); LLM evaluation frameworks (RAGAS for RAG evaluation, LangSmith for tracing, DeepEval for benchmarking); A/B testing prompt variations
• NLP pipeline development — building text preprocessing pipelines for production; tokenisation; sentence splitting; named entity recognition (NER) using spaCy or Hugging Face token classification; text classification (sentiment analysis, intent detection, topic classification); information extraction; relation extraction; coreference resolution; these classical NLP tasks remain important as pre/post-processing components around LLMs
• Conversational AI and chatbot development — intent classification and entity extraction for task-oriented dialogue systems; dialogue state tracking; multi-turn conversation management; voice-enabled NLP with ASR (Automatic Speech Recognition — Whisper, AWS Transcribe) and TTS (Text-to-Speech — ElevenLabs, AWS Polly); integration with messaging platforms (WhatsApp Business API, Facebook Messenger, web chat widgets); customer service and internal helpdesk automation
• Document intelligence and information extraction — processing unstructured documents (PDFs, scanned forms, contracts, medical reports) to extract structured information; OCR integration (Tesseract, AWS Textract, Google Document AI); layout-aware document understanding (LayoutLM, Donut); named entity extraction; table extraction; key-value pair extraction; the foundation of banking KYC automation, legal contract analysis, and government document digitisation
• Text embedding and semantic search — training and fine-tuning sentence embedding models using contrastive learning (SimCSE, supervised contrastive); evaluating embedding quality (BEIR benchmark, MTEB leaderboard); building semantic search engines using FAISS or vector databases; dense passage retrieval; bi-encoder vs cross-encoder architectures; semantic similarity and deduplication at scale
• Sinhala/Tamil NLP (Sri Lanka-specific) — the most nationally significant NLP research direction; Sinhala tokenisation challenges (agglutinative morphology, Unicode normalisation); Sinhala word embeddings (SinhalaBERT, IndicBERT); Sinhala-English code-switching NLP; low-resource language techniques (multilingual models like XLM-R, mBERT as starting points; data augmentation; cross-lingual transfer); collaboration with the Language Technology Research Lab at University of Moratuwa; a genuine first-mover opportunity
• LLM deployment and inference optimisation — serving LLMs in production; vLLM (the standard high-throughput LLM inference server using PagedAttention); TGI (Text Generation Inference by Hugging Face); quantisation (GPTQ, AWQ, GGUF for CPU/consumer GPU inference with llama.cpp); speculative decoding; batching strategies; cost optimisation (model routing: use a small cheap model for easy queries, large model for hard ones); GPU memory management
• NLP system evaluation and safety — LLM evaluation (factuality, hallucination rate, toxicity, bias); red-teaming for adversarial inputs; Constitutional AI and RLHF alignment techniques; building evaluation datasets and benchmarks; monitoring production NLP systems for drift, hallucination, and unexpected outputs; essential for deploying NLP in regulated industries (banking, healthcare)