Company Description Konnect Co. Ltd is building an AI-powered super app that simplifies how international visitors, students, and residents experience daily life in Korea. The platform removes common barriers such as language, payments, and identity requirements that typically demand a Korean phone number, national ID, or local bank account. Konnect’s mission is to serve as a digital bridge that makes Korea more accessible, welcoming, and efficient for the global community.

(https://koreatechdesk.com/k-startup-grand-challenge-ksgc-comeup-2025-winner-indias-konnect)

Job Title: Senior AI Engineer (LLM Architecture & Optimization)

Role Overview

We are seeking a Senior AI Engineer with a deep foundation in neural network architectures to lead the curation, optimization, and scaling of our language models. You will be responsible for building multilingual datasets from scratch, performing surgical modifications on transformer layer weights, implementing advanced dimension reduction on latent spaces, and deploying highly optimized quantized models into production serving engines.

Work Structure

Format: Hybrid (eventually onsite in South Korea)

Compensation: As per industry Standard - commensurate with experience

Joining: Immediate if possible

Key Responsibilities

-Neural Network Optimization: Analyze and manipulate internal neural network layer matrices ($W_q, W_k, W_v, W_{gate}$) to safely perform post-training quantization and minimize perplexity degradation.

-Dataset Engineering: Architect the end-to-end data pipeline, including sourcing, filtering, and normalizing diverse multilingual open-source datasets (e.g., Aya, Flores) to expand the model's linguistic capabilities.

- Latent Space Manipulation: Design and implement high-density dimension reduction workflows (e.g., Matryoshka Representation Learning)



on hidden states to optimize embedding pipelines without losing semantic meaning.

- High-Throughput Serving: Deploy optimized models directly into vLLM, configuring PagedAttention and continuous batching for maximum production efficiency.

- Architectural Safety: Mitigate critical neural network issues such as catastrophic forgetting, gradient interference, and activation outliers during multi-lingual fine-tuning.

Technical Requirements

- Neural Network Mechanics: Deep theoretical and practical mastery of Transformer blocks, multi-head attention mechanisms, FFN sub-layers, and backpropagation behavior.

- Inference Engines: Hands-on production experience serving models via vLLM or TensorRT-LLM.

- LLM Quantization: Proven track record with AWQ, GPTQ, BitsAndBytes, and GGUF, including a deep understanding of weight distribution and activation scaling (e.g., SwiGLU/GeLU).

- Dimension Reduction: Mastery of PCA, SVD, and modern vector quantization techniques applied directly to embedding spaces.

- Multilingual Training: Expert knowledge of SentencePiece/BPE tokenizers, cross-lingual vocabulary expansion, and managing vocabulary overlap.

- Core Ecosystem: Expert-level Python, PyTorch, and the Hugging Face library suite (transformers, peft, accelerate).

Preferred Qualifications

- Strong programming skills in C++ or CUDA for writing or modifying custom optimization kernels.
- Experience deploying high-density vector workflows in vector databases like Qdrant, Milvus, or Pinecone.
- Experience handling data crawling, deduplication, and quality filtering tools at scale.

Experience & Education

- Bachelor’s, Master’s, or PhD in Computer Science, Mathematics, Data Science, or a related quantitative field.
- 5+ years of skilled experience focusing heavily on neural networks, NLP, LLM architecture, or machine learning infrastructure.

📌 Artificial Intelligence Engineer (Bengaluru)
🏢 Konnect Co.
📍 Bengaluru