全职 Founding AI Engineer – Industrial Vision Language Models 工作, 薪水, CELOTTI TECH TALENT 公司招聘中 - Ricebowl

Celotti Tech Talent is an Australian technology recruitment firm specializing in AI, Data, Deep Tech and Robotics, with a presence across Australia and Singapore.

About Our Client

Our client is building AI systems that connect engineering design with manufacturing execution and business decisions.

At the core of their products is an industrial vision-language model that reads engineering drawings the way an experienced manufacturing engineer does understanding annotations, GD&T, materials, notes, and manufacturing features, and reasoning across views and between 2D and 3D.

Our client's mission is to turn fragmented industrial information into actionable intelligence and deliver measurable improvements in engineering productivity, manufacturing quality, and operational efficiency.

Position Overview

Hiring Founding AI Engineers to build and own industrial vision-language models (VLMs).

You will work directly with our founders, Chief Scientist, and manufacturing domain experts to turn open multimodal models into a model that reliably understands engineering drawings and CAD and to prove it with rigorous benchmarks on real customer data.

This is a hands-on model-development role spanning data, training, evaluation, and serving. You will own the model end to end, not the application layer around it.

You will have access to a large GPU cluster with no budget cap on compute model quality, not hardware, is the constraint.

What You’ll Build

Drawing & CAD Understanding Model

• Adapt and train VLMs to read 2D engineering drawings: dimensions, tolerances, GD&T frames, surface finish, title blocks, bills of materials, notes, and section and detail views.
• Extend the model to multi-view and 2D–3D reasoning: linking drawing views to 3D CAD/STEP geometry and recognizing manufacturing features such as holes, threads, pockets, fillets, and chamfers.
• Train the model to call tools - OCR, symbol detectors, geometry and CAD kernels where they are more reliable than pixels alone.

Industrial Data Engine

• Build pipelines to collect, clean, and de-identify industrial drawings and CAD models, and to generate synthetic drawings from CAD at scale.
• Design annotation schemas and semi-automated labeling tools with engineer-in-the-loop review.
• Turn engineer corrections from production use into new training and evaluation data.

Training & post-training

• Run continued pretraining, supervised fine-tuning, and preference or RL-based post-training (e.g., DPO, GRPO) for high-resolution, text-dense technical images.
• Make deliberate choices on vision encoders, image resolution and tiling, and parameter-efficient versus full fine-tuning.
• Run large-scale, multi-node training efficiently and keep experiments reproducible and well documented.

Evaluation & Deployment

• Build benchmarks for field-level extraction accuracy, GD&T and tolerance interpretation, feature recognition, and cross-view consistency, with error analysis by drawing type and failure mode.
• Calibrate model confidence so uncertain outputs are flagged for engineer review instead of guessed.
• Serve models efficiently (e.g., vLLM, quantization) and track performance on real customer drawings.

Key Responsibilities

• Own the industrial VLM end to end from data and training recipes to benchmarks, deployment, and iteration.
• Shape the multimodal model roadmap with the Chief Scientist, balancing model quality, research bets, and customer timelines.
• Decide when to train, when to fine-tune, and when to combine the model with OCR, geometry processing, or rules.
• Report results with clear metrics and honest failure analysis.
• Work with manufacturing engineers and customers to define what “correct” means on real drawings.
• Publish or open-source selected work where it strengthens our technical position.

Qualifications

Required

• B.S., M.S., or Ph.D. in Computer Science, AI, Electrical Engineering, Mechanical Engineering, or a related field—or equivalent hands-on experience.
• Hands-on experience training or fine-tuning vision-language or multimodal models not only prompting or retrieval with results you can show.
• Strong Python and PyTorch skills, including multi-GPU or distributed training (e.g., DeepSpeed, FSDP) and the Hugging Face ecosystem.
• Experience building multimodal training and evaluation datasets and improving models through systematic error analysis.
• A builder’s mindset: strong ownership, comfort with ambiguity, and drive to make models work on real industrial data.

Preferred

• Document AI, OCR, or understanding of text-dense images such as tables, forms, or technical drawings.
• Experience with engineering drawings, GD&T, STEP/B-Rep, CAD APIs, or 3D geometry learning.
• Post-training with preference optimization or reinforcement learning; synthetic data generation.
• Large-scale, multi-node training and inference optimization (e.g., vLLM, quantization, high-resolution or long-context serving).
• Publications at venues such as CVPR, ICCV, ECCV, NeurIPS, ICML, ICLR, ACL, EMNLP, or ACM MM, or significant open-source contributions.
• Familiarity with manufacturing processes, DFM, or industrial standards.

Why Join?

• Build the company’s core model as a member of the founding engineering team.
• Train on real industrial drawings and CAD from manufacturing customers data most AI teams never see.
• Train at scale on a large GPU cluster, with no budget cap on compute.
• Work directly with the Chief Scientist and manufacturing domain experts on technically challenging problems.
• Receive competitive salary, meaningful equity, and significant opportunities for professional growth.
• Singapore Employment Pass sponsorship support for eligible candidates.