Engineering Notes · Go · GenAI · Infrastructure
How Claude + Gemini Gave Me a Crash Course in Go, H3 Hexagons, and Cloud Infrastructure — In Under 3 Hours
Thushara Wijeratna · June 2026
There’s a moment in any engineering project where you hit a wall that isn’t really a technical wall — it’s a knowledge wall. You know what you want to build. You can see the shape of the solution. But the specific language, library, or paradigm you need is outside your comfort zone, and the time cost of getting up to speed feels prohibitive.
I hit exactly that wall recently. And I got through it in about two and a half hours, with a working solution, a complete deployment pipeline, and a meaningful education. Here’s how.
The Background: A Cluster Migration and an Unexpected Opportunity
We’re in the middle of migrating a geo-processing cluster from Rails to Go. The new cluster is more efficient by an order of magnitude, but it’s barely past proof-of-concept — just enough tested to know it works, not enough to bet production on. It runs on ECS and has access to our Redis Time Series infrastructure.
Then a separate project came up: mapping driver density into H3 hexagons. The idea is to take raw GPS driving data and visualize where drivers are concentrated — a spatial density problem. It’s the kind of analysis that screams for Python or Go. Not Ruby. Definitely not Rails.
The problem: I barely understand Go. And while Python is friendlier territory, the natural home for this work was the new Go cluster I was already running — it had the right access to the time series data, and using it here would be a meaningful first real exercise for the cluster.
So I had a choice: spend days self-teaching Go spatial libraries, Docker packaging, ECS task definitions, and IAM policies — or ask for help. I asked for help.
Phase 1: Claude Builds the Foundation
I opened a conversation with Claude and described what I was after: take a sample of driving data, bin the GPS coordinates into H3 hexagons at an appropriate resolution, and produce a density visualization. I explained the Go context and asked for a working implementation.
What I got back wasn’t just code. It was reasoned code. Claude walked through the H3 resolution tradeoffs (too fine and you get sparse, noisy hexagons; too coarse and you lose geographic meaning), chose a sensible default, and produced a complete Go program that:
- Parsed the GPS data
- Mapped each coordinate to its H3 cell index
- Counted occurrences per hexagon
- Exported the density data ready for visualization
The code compiled. The logic made sense when I read it. When I hit small integration issues, Claude adjusted immediately when I fed back the errors.
But the analysis code was only part of the problem.
Phase 1b: The Infrastructure Claude Also Built
This is the part of the story I want to make sure doesn’t get lost — because it’s where the time savings were arguably even more dramatic.
To run this Go script in the new cluster, I needed a complete infrastructure stack built from scratch:
A Dockerized Go application. The script had to be packaged as a container, with the right base image, dependency management, and build configuration to run cleanly inside ECS. Claude produced a working Dockerfile and explained each layer.
An ECS Task Definition. The container needed to be provisioned as a task in the new geo cluster — with the right CPU/memory allocation, environment variables, logging configuration, and networking setup to reach the Redis Time Series instance. This meant understanding how ECS task definitions are structured and how they interact with the surrounding infrastructure.
IAM roles and policies. The tricky part: the driving data lives in a locked-down production environment. To process it safely, the task needed read access to an S3 bucket containing that data, without being given any broader permissions it didn’t need. Claude helped design the IAM policy with least-privilege principles — specific bucket ARNs, specific actions, nothing more. It also helped reason through the trust relationship between the ECS task execution role and the task role itself, which is a subtle distinction that trips up a lot of people.
S3 bucket configuration. A staging bucket was needed to land the processed data safely — isolated from production, with appropriate access controls so the analysis job could write results without touching anything it shouldn’t.
What I got wasn’t just “here’s a Dockerfile.” It was a coherent infrastructure design: here’s how the pieces connect, here’s why we’re scoping the IAM policy this way, here’s what would go wrong if we didn’t. The guardrails weren’t an afterthought — they were built in from the start.
By the end of that session, I had a traffic density plot from real driving data, running inside the new cluster, reading from production data safely through a well-scoped IAM boundary, writing results to an isolated S3 bucket. Working. Deployed. Safe.
If the story ended there, it would already be remarkable. A non-Go developer, unfamiliar with ECS task definitions and spatial indexing libraries, with a working, deployed pipeline in a couple of hours. That’s the headline.
But the story didn’t end there.
Kepler.gl rendering of H3 resolution-5 driver density over the Dallas–Fort Worth metro. The yellow hexagon (Garland/East Dallas) logged 1,856,072 trips in the sample — the output of the Go pipeline built in this session.
Phase 2: Gemini Pressure-Tests and Extends the Vision
I took the Claude-generated solution to Gemini — not to debug it, but to challenge it. I wanted to know: is this approach actually good? And what would make it better?
This is a workflow pattern worth naming explicitly: use one AI to build, use another to audit and extend. Different models have different strengths, different training emphases, different tendencies in how they approach problems. Running your solution past a second model is a cheap form of peer review.
Gemini’s response surprised me. It didn’t just validate the approach — it pushed it forward. The suggestions fell into two categories:
1. Statistical enrichment at scale. The density map I’d built was a point-in-time snapshot. Gemini pointed out that with more data, you could extract genuinely useful statistics: density gradients, hotspot persistence over time, anomaly detection for unusual clustering patterns. These aren’t just nice-to-haves — for a driver dispatch system, they’re operational intelligence.
2. Redis Time Series as the efficient backbone. Here’s where it got interesting. Gemini connected the density problem to infrastructure I already had running in the cluster: Redis Time Series. With compaction rules, you can store high-frequency GPS events efficiently and roll them up into time-windowed density summaries without blowing out memory or query latency. The density map becomes a living view into an efficiently maintained time series, not a batch job.
This wasn’t something I would have reached on my own in a single session. It required knowing that Redis Time Series could do this, understanding how compaction rules work, and seeing the connection between spatial density and temporal aggregation. Gemini handed me all three in one exchange.
What I Actually Learned
The output of this session wasn’t just a working program. It was an accelerated education across several domains I wasn’t strong in:
Go development — not just syntax, but idiomatic patterns for data processing pipelines, error handling, and module organization. Reading and debugging Claude’s code taught me more about Go in two hours than I’d absorbed from documentation in weeks.
H3 spatial indexing — the geometry of hexagonal hierarchical indexing, resolution tradeoffs, and why H3 is better suited for density mapping than geohash or simple grid approaches.
ECS + IAM design — how to structure a task definition, how to reason about least-privilege IAM policies for ECS workloads, and how to isolate production data access safely. This alone would have taken a full day to research and implement from scratch.
Stats at scale — how to think about density not as a static count but as a statistical distribution with temporal dimensions, and how to maintain that data efficiently in a system you’re already running.
This is the thing I keep coming back to: Gen AI isn’t just a code generator. Used well, it’s a compressed curriculum. You get the code, but you also get the reasoning, the tradeoffs, the “here’s why this approach and not that one.” If you pay attention, you learn.
The Workflow, Distilled
1. Bring a concrete problem, not an abstract question. “Build me a density map of GPS data in Go using H3, packaged for ECS, reading from S3 with a scoped IAM role” is better than “how do I do spatial analysis in Go.” Specificity gets specific answers.
2. Don’t just run the code — read it. The code is the lesson. When something compiles and runs, go back and understand why each piece is there. Ask the AI to explain sections that aren’t clear. This is where the actual learning happens.
3. Use a second model as a peer reviewer. After you have something working, bring it to a different model and ask: “Is this approach sound? What am I missing? How would this scale?” You’ll get different angles, different critiques, different extensions.
4. Follow the threads. Gemini’s Redis Time Series suggestion wasn’t something I asked for. It emerged from a conversation about making the approach more robust. Let the conversation go where it wants to go — some of the best insights come from tangents.
A Note on What This Isn’t
This isn’t a story about AI replacing engineering judgment. The H3 resolution I used, the decision to pursue Redis Time Series compaction, the choice of which statistics actually matter for driver dispatch, the architecture of which IAM boundaries made sense given our production setup — those required domain understanding that I brought to the table. The AI provided implementation and expanded my knowledge of available tools. The judgment about what to do with those tools remained mine.
That division of labor — AI handles implementation and education, engineer handles domain judgment and architectural decisions — is the productive pattern. Collapse it in either direction and you lose something important.
Closing Thought
Two and a half hours. One working density visualization. A complete Docker + ECS + IAM + S3 deployment pipeline. A crash course in Go, spatial indexing, cloud infrastructure, and time-series-backed analytics. A roadmap for how to make the analysis genuinely production-grade.
The combination of Claude for building and Gemini for pressure-testing and extending isn’t a fluke. It’s a repeatable workflow for the class of problems where you know what you want but don’t yet have the specific technical fluency to get there alone.
The 10x output claim isn’t about volume. It’s about the ratio of what you can accomplish to what your current knowledge would otherwise limit you to. That ratio, in the right problem, can be extraordinary.
The Claude transcript showing how the Go/H3 solution and infrastructure were built is available here. The Gemini conversation on statistical extensions and Redis Time Series is here.
No comments:
Post a Comment