Use Case: Spatial Data Scientist in Carbon Accounting at Amazon Climate Tech

Most candidates burn out on SQL screens. The ones who fail at Amazon's Climate Tech org fail on ownership. They model carbon flux in hectares of Indonesian peatland and cannot explain who should act when their model flags a supplier at 12,000 tons CO2e over contract. In a December 2023 debrief for the Spatial Data Scientist role on Amazon's Carbon Accounting team, the hiring manager—a Principal PM who had spent four years at the World Resources Institute before joining Amazon—voted no-hire on a candidate from Planet Labs. Candidate had built a beautiful time-series model for deforestation in the Congo Basin.

Beautiful. Useless. The model output sat in a dashboard nobody monitored. The problem wasn't the answer. It was the judgment signal. The candidate never asked who paid when the model was wrong.

----

What Does a Spatial Data Scientist Actually Do in Amazon Climate Tech?

You build attribution models that connect satellite pixels to supply chain decisions. Not research. Operations.

The Carbon Accounting team sits inside Amazon's Sustainability organization, reporting up through the Global Vice President of Sustainability. In 2022, Amazon restructured this team after failing its 2040 net-zero commitment timeline in an internal audit. The Spatial Data Science function was created specifically to solve the "last mile" problem of carbon accounting: suppliers claiming sustainable sourcing in regions where ground verification was impossible. Your job is to make the unverifiable, verifiable. Then make it actionable.

I sat in a Q2 2024 hiring committee for a level 6 position on this team. The role description mentioned "remote sensing expertise" and "machine learning." What the hiring manager—who had joined from Stripe's Climate team in 2023—actually wanted: someone who could build a deforestation-risk score for Brazilian soy suppliers that the Amazon procurement team would actually use. Not a paper. A score. Integrated into Oracle supplier management workflows. Updated quarterly. Audited annually by Ernst & Young for CDP disclosure.

The team uses a specific stack. Google Earth Engine for satellite data ingestion. Python (xarray, rasterio) for processing. AWS SageMaker for model training. But the critical tool is not technical.

It is the "Working Backwards" document that every Spatial Data Scientist must write before modeling. This six-pager forces you to define the customer—often an Amazon Senior Vendor Manager who spends 11 hours a day in vendor negotiations in Shenzhen—and their decision before touching a pixel. In a February 2024 debrief, a candidate from NASA JPL spent 20 minutes on atmospheric correction algorithms. The Amazon Principal Scientist interrupted: "Who is the customer and what do they do differently after your model?" The candidate answered, "They have better information." No hire. Unanimous.

Compensation for this role at L6 runs approximately $182,000 base, 0.04% equity vesting over four years, and $35,000 sign-on. L7 adds RSU refreshers and a broader scope across multiple commodity categories. The equity percentage sounds small. At Amazon's market cap, 0.04% is material. The sign-on is negotiable within a band; I have seen $28,000 and $42,000 for candidates with competing offers from Microsoft Climate or Google Sustainability.

----

How Is the Interview Structured for This Amazon Role?

Five loops. One of them will be a bar-raiser you do not expect. The bar-raiser in my December 2023 debrief was from Alexa Shopping. Knew nothing about carbon. Cared deeply about whether the candidate understood second-order effects.

The structure: (1) Phone screen with hiring manager, 45 minutes, (2) Technical screen with senior data scientist, 60 minutes, (3) Onsite loop with five interviews, 45 minutes each, (4) Debrief, (5) Hiring committee review. Timeline from application to offer: 47 days average in my experience, though Q4 2023 stretched to 71 days due to reorg freezes.

The technical screen is not a LeetCode exercise. In a 2023 loop for this exact role, the candidate was given Planet NICFI basemaps of Rondonia, Brazil, from 2019 and 2023. Task: calculate forest loss, attribute to likely commodity conversion, estimate carbon emissions using IPCC Tier 1 factors, and present a one-slide recommendation to a hypothetical VP.

The candidate who passed—now an L6 on the team—spent the first 10 minutes asking about the VP's existing supplier scorecard, not the imagery. "I needed to know if I was building a new signal or replacing a broken one," she told me later. That is the judgment signal Amazon tests for.

The onsite loop includes: (1) Data Science Deep Dive, (2) Leadership Principles (two rounds, always), (3) Career History, (4) Bar Raiser, (5) "Ownership" case study specific to Climate Tech. The case study in my Q2 2024 debrief involved a palm oil supplier in Malaysia whose satellite-verified deforestation exceeded the "No Deforestation, No Peat, No Exploitation" commitment in their supply contract. The candidate's task: model the emissions impact and recommend action. The candidate who advanced to offer spent 15 minutes on the model.

The remaining 30 on stakeholder negotiation: "The supplier claims the clearing was for a local hospital. My model says 847 hectares. How do I verify? Who do I escalate to? What if Legal says we cannot terminate without physical audit?"

The Leadership Principles are not decorative. "Insist on the Highest Standards" and "Dive Deep" appear in every Climate Tech loop I have reviewed. The "Dive Deep" question in a 2023 loop: "Tell me about a time your model produced an unexpected result and you had to determine if it was a bug or a real signal." The candidate who passed described a three-week investigation into Sentinel-2 band misalignment that turned out to be a real phenological anomaly.

The candidate who failed described a "quick check" and immediate escalation. Surface. No ownership.

----

> 📖 Related: Amazon RTO Interview vs Google Hybrid: Whiteboard Depth vs Virtual Flexibility

What Technical Skills Actually Matter for This Role?

Not deep learning. Not computer vision. Spatial statistics at scale with operational constraints.

The Carbon Accounting team has tried deep learning. In 2022, they ran a six-month pilot with convolutional neural networks for land cover classification in the Cerrado. Accuracy was 4% higher than random forest. Inference cost was 340% higher. Latency made real-time supplier scoring impossible. The project was killed in a Q1 2023 portfolio review. The lesson, now embedded in the hiring rubric: elegant solutions that do not operate at Amazon's scale and speed are not solutions. They are hobbies.

What the team actually uses: Random Forest and Gradient Boosting (XGBoost, LightGBM) for classification. Bayesian models for uncertainty quantification. Time-series decomposition (STL, Prophet) for trend detection in forest disturbance. All models must run on AWS infrastructure costing under $0.02 per hectare analyzed, a constraint added to the role after the 2022 restructuring.

The critical skill is not model selection. It is geospatial data engineering at planetary scale. Ingesting MODIS, Landsat, Sentinel-1, and Sentinel-2. Harmonizing spatial resolutions. Handling cloud masking across sensors.

The candidate from Planet Labs who failed in December 2023? His models were excellent. His data pipelines were manual. "I would need a data engineer to productionize," he said in the loop. At Amazon, you are the data engineer. The job posting says "Spatial Data Scientist." The reality is "Spatial Data Scientist who owns the full stack from satellite downlink to executive dashboard."

Specific tools tested: Google Earth Engine (mandatory), Python (rasterio, xarray, geopandas), SQL (Redshift for supply chain joins), and AWS (SageMaker, Lambda, S3). Tableau or QuickSight for visualization. Not optional: experience with carbon accounting frameworks. GHG Protocol. SBTi.

ICVCM Core Carbon Principles. In a 2024 loop, a candidate was asked to explain the difference between Scope 3 Category 1 (Purchased Goods and Services) and Category 4 (Upstream Transportation and Distribution) emissions in the context of soy supply chains. She paused for 12 seconds. Then answered with specific emission factors from EXIOBASE and the limitations of Ecoinvent for agricultural commodities. Hired at L7.

Compensation negotiation for technical skills: candidates with production experience in MRV (Measurement, Reporting, Verification) for voluntary carbon markets command 15-20% base salary premiums. I have seen $198,000 base for a candidate with Verra VCS methodology development experience. The market is thin. Amazon knows it.

----

How Do You Demonstrate Ownership in Carbon Accounting Interviews?

Ownership at Amazon Climate Tech means defining the problem before solving it, then living with the consequences.

In a Q3 2023 debrief, the hiring committee was split 3-2 on a candidate from The Nature Conservancy. Strong technical skills. Weak ownership signal. The deciding factor: a question about a time the candidate disagreed with a stakeholder.

The candidate described a situation where her regional director wanted to publish a deforestation analysis before ground-truthing. "I explained the risks and he agreed to wait." The bar-raiser pushed: "What was the cost of waiting?" The candidate had not calculated it. Did not know the supplier contract renewal date. Had not modeled the reputational risk of delayed disclosure versus inaccurate disclosure. No hire.

The candidate who was hired in that same loop—now leading the team's Southeast Asia portfolio—told a different story. He had identified that Amazon's existing carbon accounting methodology double-counted emissions from smallholder farms in Indonesian palm oil supply chains. The double-counting inflated reported emissions by 8%. He did not simply report the bug.

He built a corrected methodology, socialized it with the Science Based Targets initiative verification team, implemented the fix in Amazon's ERP system, and presented the change to the CDP disclosure committee. Total timeline: seven months. The "Ownership" signal was not finding the error. It was the seven months of institutional negotiation to fix it.

Script for the "Ownership" Leadership Principle, as validated by successful candidates in this loop:

Not: "I identified a problem and fixed it."

But: "In Q2 2023, I discovered our MODIS-based forest loss alert was triggering 14 days after physical clearing. For a supplier compliance use case, 14 days meant the violation was already in our supply chain. I benchmarked against Global Forest Watch's GLAD alerts, prototyped a Sentinel-1 SAR integration reducing latency to 36 hours, convinced the data engineering team to reprioritize their sprint, and we cut response time by 89%. The supplier termination decision now happens before shipment, not after."

The specific numbers matter. The stakeholder names matter. The timeline matters. Generic ownership stories fail because they read as fabricated. Specific stories with real product names and quantified outcomes pass because they demonstrate the judgment Amazon pays for.

----

> 📖 Related: Google L4 vs Amazon L5 Total Comp for PMs in 2025

Preparation Checklist

  • Build one end-to-end project from satellite imagery to business decision using real data (the PM Interview Playbook covers Amazon-style "Working Backwards" documents for technical roles with actual debrief examples from Climate Tech loops)
  • Practice explaining cloud masking, atmospheric correction, and spatial resampling to a non-technical audience—specifically, practice explaining to someone from Finance or Legal
  • Complete at least one carbon accounting exercise using GHG Protocol Corporate Standard methodology, not just academic LCA
  • Set up a personal AWS account and run a geospatial pipeline on real satellite data; production-like constraints beat academic accuracy
  • Prepare three specific "Ownership" stories with named stakeholders, dollar impacts or emission quantities, and timelines exceeding six months
  • Review Amazon's 2023 Sustainability Report and identify one specific gap where spatial data could improve disclosure accuracy

----

Mistakes to Avoid

BAD: "I built a machine learning model to predict deforestation with 94% accuracy."

GOOD: "I built a deforestation-risk score for soy suppliers in the Cerrado that the procurement team used to renegotiate contracts covering 2.3M tons annually, with a false positive rate below 5% to avoid supplier disputes."

BAD: "I am passionate about climate change and want to use my skills for good."

GOOD: "At my current role, I identified that our Scope 3 methodology was using outdated emission factors for Indonesian peatland. I updated the factors, recalculated three years of baselines, and the revised numbers were accepted by our external auditor."

BAD: "I would present the findings to the relevant stakeholders."

GOOD: "I drafted the memo, scheduled the decision meeting with the VP of Sustainability, pre-briefed the two dissenting directors, and had the supplier termination letter ready for Legal review before the meeting."

Pitfall 1: Over-indexing on model accuracy. In a 2023 loop, a candidate from a top-5 PhD program spent 18 minutes on F1 score optimization. The Amazon Principal Scientist stopped him: "Your model is 3% more accurate and costs $40,000 more per run. The procurement team has a $2M annual tooling budget. Convince them." The candidate could not. No hire.

Pitfall 2: Ignoring the political economy of supply chains. Carbon accounting does not exist in a vacuum. In a 2024 debrief, a candidate proposed suspending a major Brazilian meatpacker based on satellite evidence of illegal deforestation. She had not considered that the meatpacker supplied 23% of Amazon's fresh meat in Brazil, that suspension would require invoicing complications, or that the supplier held leverage in an ongoing price renegotiation. The hiring manager: "Right conclusion, wrong process. You can't just be correct. You have to be implementable."

Pitfall 3: Treating Amazon like an research institution. The December 2023 Planet Labs candidate asked about publication policies. "Can we publish on this methodology?" The hiring manager's face, as described to me by the recruiter: "like he'd asked about vacation days in the first five minutes." Amazon Climate Tech publishes sparingly. The work is proprietary competitive advantage. Candidates who signal academic career intentions are screened out regardless of technical strength.

----

FAQ

How long does the Amazon Spatial Data Scientist interview process take from application to offer?

47 days average, 71 days in reorg periods. The Carbon Accounting team runs quarterly hiring waves tied to AWS budget cycles. Apply in Q1 or Q3 for fastest turnaround. Q4 is frozen for most external hiring. The phone screen-to-onsite ratio is tight; I have seen 3:1 in this specialized role. One candidate in 2023 waited 19 days between onsite and debrief because the bar-raiser was on vacation. Amazon does not substitute bar-raisers. Ever.

What is the typical compensation for Spatial Data Scientist at Amazon Climate Tech?

L6: $182,000 base, 0.04% equity, $28,000-$42,000 sign-on. L7: $210,000-$235,000 base, 0.06%-0.08% equity, larger sign-on. The equity is front-weighted in years 3-4 to improve retention. In 2023, Amazon lost two Spatial Data Scientists to Stripe Climate at 30% premium; compensation bands have widened since. Negotiate with specific competing numbers, not vague market data. The hiring manager has limited flexibility on base but can advocate for sign-on and equity within band.

How technical is the "technical" interview, and what should I prioritize?

Not technical enough to justify PhD-level methodology. Technical enough to expose gaps in production thinking. Prioritize: data pipeline architecture over model architecture, operational constraints over theoretical elegance, stakeholder management over statistical significance. In my last five debriefs for this role, the successful candidate was not the one with the most sophisticated model. It was the one who asked, after 10 minutes of problem-solving, "What is the cost of being wrong, and who pays?"amazon.com/dp/B0GWWJQ2S3).

Related Reading

What Does a Spatial Data Scientist Actually Do in Amazon Climate Tech?