New opportunity

Machine Learning Engineer

Unravel Carbon Pte Ltd · Singapore

About this role

Unravel Carbon seeks an early-career Machine Learning Engineer focused on generative AI to build production systems for its sustainability platform. The role works with product managers, software engineers, designers, data engineers and sustainability experts, and offers an opportunity to shape product direction and the ML team’s culture.

Responsibilities

Build AI agents that automate emissions calculations, extract insights from emissions inventories, generate sustainability reports and streamline other workflows.
Develop LLM-powered systems to process invoices, utility bills and supply chain documents. Build agentic workflows combining language models, structured data processing and sustainability-specific reasoning.
Improve a carbon-emissions engine using semantic search and text classification across millions of transaction records and a proprietary emissions-factor dataset.
Build scalable, modular ML services on AWS for production LLM systems and transformer-based models, including ModernBERT and T5. Implement LLM evaluation frameworks and quality controls, and develop systems that recommend decarbonization solutions.

Requirements

The posting calls for 2–5 years building production ML systems, including hands-on work with LLMs or other generative models; strong recent graduates with substantial relevant internship or research experience may also apply.
Strong Python and SQL skills; familiarity with transformers, LangChain or similar libraries, prompt engineering, and LLM APIs or RAG systems. Applicants should have shipped impactful NLP or ML solutions to production.
Ability to own the ML project lifecycle, design scalable services, write maintainable code, use AI-based coding tools effectively, collaborate across functions and adapt to changing priorities. The role also values mentoring.

Preferred qualifications include LLM evaluation, guardrails or agent-framework experience; vector databases and embeddings; sustainability, carbon accounting or climate-tech knowledge; and a master’s or PhD in Computer Science, Applied Computing, Computational Linguistics, Machine Learning or a related field. The position is listed in Singapore and offers work with founders and senior leadership on production AI systems for sustainability.

For full details about this role, click “Apply on company site” to view the employer’s original posting.

Skills for this role

PythonSQLMachine LearningGenerative AILarge Language ModelsAI AgentsNatural Language ProcessingPrompt EngineeringRetrieval-Augmented GenerationLangChainTransformersModernBERTT5Semantic SearchText ClassificationAWSML System DesignLLM EvaluationAI Coding ToolsCross-functional CollaborationMentoringVector DatabasesEmbeddingsCarbon Accounting

Checking your profile…

Preparation suggestions, not the employer’s actual interview questions. Always check the original posting for current requirements.

Understand the role

What matters most

Build sustainability AI agents
The role builds agents for emissions calculations, inventory insights and sustainability reports, including workflows that combine LLMs, structured data and domain reasoning.
Process sustainability documents
The posting specifically calls for LLM systems that understand invoices, utility bills and supply-chain documents. Prepare to explain how document data becomes usable structured information.
Improve the carbon-emissions engine
A core duty is improving semantic search and text classification over millions of transaction records and a proprietary emissions-factor dataset.
Ship reliable ML services
The role designs scalable, modular AWS services for transformer models and LLM systems, and implements LLM evaluation frameworks and production quality controls.
Turn ML work into customer value
The engineer will help build decarbonization recommendations and work with product, engineering, design, data and sustainability colleagues. The posting also expects ownership across the ML project lifecycle.

Likely challenges to think through

Automation versus emissions accuracy
Likely preparation challenge, not a confirmed problem: agents that calculate emissions or generate reports must be useful without treating uncertain document interpretations as facts. Prepare to discuss validation, traceability and when to seek human review.
Model quality versus production scale
Likely preparation challenge: processing millions of transactions while serving both LLM and transformer systems may require balancing accuracy, latency and maintainability. Prepare a production ML example that explains those trade-offs and how quality was measured.
Product speed versus domain understanding
Likely preparation challenge: shaping workflows and decarbonization recommendations across disciplines may require moving quickly while checking sustainability-specific assumptions. Prepare to explain how you would clarify the user problem and verify domain reasoning with relevant colleagues.

Practice your answers

Technical practice

How would you extract usable data from invoices, utility bills, and supply chain documents?

What to coverThe posting calls for LLM-powered processing of these document types. Suggested preparation: discuss extraction into a structured schema, validation of ambiguous fields, and how you would measure accuracy before using the data in emissions calculations.

How would you design an agent that calculates emissions and drafts a sustainability report?

What to coverThe role includes agents for emissions calculations, inventory insights, and reports. Suggested preparation: explain which steps need structured calculations rather than generated text, and how you would make inputs and outputs reviewable.

When would you use semantic search versus text classification in the carbon engine?

What to coverThe posting says the engine uses both methods across millions of transaction records and a proprietary emissions-factor dataset. Suggested preparation: compare their roles in finding candidate matches and assigning categories, including accuracy and scale trade-offs.

How would you investigate incorrect emissions-factor matches?

What to coverThe posting calls for continuous improvement of an engine that calculates emissions from transactional data. Suggested preparation: trace a result from source transaction through model output to factor selection, then describe how you would diagnose errors and verify a fix.

How would you structure ML services on AWS for both transformer models and LLM systems?

What to coverThe role calls for scalable, modular production services on AWS, including models such as ModernBERT and T5. Suggested preparation: discuss service boundaries, deployment and monitoring choices, and trade-offs among latency, cost, and maintainability without assuming a particular AWS product.

How would you evaluate an LLM workflow before and after release?

What to coverThe posting explicitly calls for evaluation frameworks and production quality controls; guardrails are listed as a bonus. Suggested preparation: describe representative test cases, checks for factual and calculation errors, and how you would detect regressions in production.

How would you decide whether an LLM workflow needs retrieval?

What to coverThe posting mentions LLM APIs or RAG experience and work with sustainability data from multiple sources. Suggested preparation: compare retrieving relevant source material with relying on a prompt alone, addressing evidence quality, changing data, and failure cases.

How would you choose between a transformer-based model and an LLM for a text-processing task?

What to coverThe role serves transformer-based models such as ModernBERT and T5 alongside LLM-based systems. Suggested preparation: explain how task type, labeled data, reliability, latency, and cost would inform the choice rather than assuming one model fits every task.

How would you build a maintainable Python and SQL pipeline for transaction-based ML?

What to coverThe posting asks for strong Python and SQL skills, maintainable code, and work at the scale of millions of transaction records. Suggested preparation: cover data checks, reproducible transformations, testing, and efficient querying as design considerations.

How would you evaluate AI-generated decarbonization recommendations?

What to coverThe role includes building systems that use company datasets to identify and recommend decarbonization solutions. Suggested preparation: discuss how you would check whether a recommendation is supported by the available data, relevant to the customer, and suitable for human review.

Behavioural practice

Tell me about a time you worked across functions to define an ML product problem.

What to coverThe posting describes collaboration with product managers, engineers, designers, data engineers, and sustainability experts. Suggested preparation: use a real situation, explain how you reconciled different needs and what you did, then state the resulting decision or outcome.

Tell me about a time you owned an ML project through production.

What to coverThe role calls for ownership of the full ML project lifecycle and evidence of production impact. Suggested preparation: describe the initial problem, your decisions from development through release, and the measured result or lesson learned.

Tell me about a time priorities changed during a technical project.

What to coverThe posting values adapting quickly in a fast-paced environment. Suggested preparation: describe what changed, how you adjusted scope or sequencing and communicated it, and what happened as a result.

Tell me about a time you raised a quality concern candidly and constructively.

What to coverThe team emphasizes intellectual honesty, transparency, reliable systems, and a strong writing culture. Suggested preparation: give a real example of the concern, how you presented evidence and proposed action, and the eventual outcome.

Tell me about a time you helped another engineer improve their work.

What to coverThe posting says the team values mentoring and maintainable code. Suggested preparation: describe the situation, the specific support or feedback you provided, and what changed for the work or the person.

Sharpen your knowledge

Topics to prepare

Agentic sustainability workflows
Suggested revision exercise: diagram how an agent could extract data from a utility bill, combine it with structured data, calculate emissions and draft a report; mark where human review may be needed. Relevant because the role builds agents that combine language models, structured processing and sustainability-specific reasoning.
Semantic search and text classification
Suggested revision exercise: compare how retrieval and classification could each help process a transaction description, including likely error cases. Relevant because both are used in the posting’s carbon-emissions engine across millions of transaction records and an emissions-factor dataset.
LLM reliability and evaluation
Suggested revision exercise: define test cases and success criteria for extracting fields from invoices and generating sustainability reports, including missing or unsupported information. Relevant because the role implements LLM evaluation frameworks and production quality controls.
Production ML lifecycle and service design
Suggested revision exercise: walk through the decisions from defining a product problem to deploying, measuring and maintaining a modular ML service. Relevant because the posting expects project-lifecycle ownership and scalable services on AWS for transformer-based and LLM systems.
Sustainability and carbon-accounting context
Suggested revision exercise: explain, at a high level, how transaction data, emissions factors, emissions inventories and decarbonization recommendations relate; flag assumptions you would check with a sustainability expert. Domain knowledge is a bonus qualification, while these data and workflows are central to the advertised work.

Skills to revise

Python and SQL data processing
Suggested practice: write Python code and SQL queries to clean sample transaction records, join them to a small emissions-factor table and check unmatched records. Both languages are explicit requirements, and the carbon engine uses transactional and emissions-factor data.
LLM document workflow
Suggested practice: prototype a prompt and LLM API workflow that extracts structured fields from sample invoices or utility bills, then checks for absent fields. The role calls for prompt engineering and LLM API or RAG experience to process these documents.
Transformer-based retrieval or classification
Suggested practice: build a small semantic-search or text-classification example using a transformer library, and inspect incorrect matches. The posting names these methods for its carbon engine and cites transformer-based models such as ModernBERT and T5.
Modular ML service implementation
Suggested practice: sketch and code separable interfaces for data input, model inference and output validation, then explain how the components could be served on AWS without assuming a particular AWS service. Relevant to the stated need for scalable, maintainable production ML services.
Evaluation and quality checks
Suggested practice: create a small test set for document extraction or report generation, score output quality and record failure categories. This directly rehearses the posting’s responsibility for LLM evaluation frameworks and production quality controls.

Your preparation checklist

Map your evidence to the role
Prepare a one-page inventory of relevant Python, SQL, production ML, LLM or generative-AI, and transformer work. Note your specific contributions and measurable results where available; the posting asks for demonstrated production impact.
Rehearse a production project walkthrough
Choose one NLP or ML project and prepare a concise account of the problem, your decisions across the project lifecycle, deployment, maintainability, results, and what you would change.
Sketch a document-to-emissions workflow
As a practice exercise, diagram how an agent might process invoices or utility bills, combine extracted information with structured data, and support emissions calculations. Mark where domain reasoning or human review would be needed.
Practice a carbon-engine design exercise
Outline how semantic search and text classification could help match large volumes of transactions to emissions factors. Prepare to explain data preparation, Python and SQL use, model trade-offs, and how you would check match quality.
Draft a production service design
Sketch a modular AWS-based service for an LLM workflow alongside a transformer model such as ModernBERT or T5. Be ready to discuss scaling, interfaces, failure handling, and code quality without assuming a particular stack beyond the posting.
Build an evaluation plan
Prepare a small example test set and proposed checks for document extraction or report generation, including error categories, reliability measures, and quality controls. Label guardrails as a potential extension; the posting lists them as a bonus.
Prepare collaboration examples and questions
Write brief examples of working with product, engineering, or domain colleagues, adapting priorities, and mentoring if applicable. Bring questions about how the team evaluates production agents, works with sustainability experts, and prioritizes product direction.

Availability last checked: 4 Nov 2026. Status can change after a check; confirm on the employer’s website. Report an issue · Our checking policy

KEEP EXPLORING

Similar AI jobs

Related skills and specializations in Singapore. Matched to this role, not your profile.

Explore more jobs

Finding similar opportunities…