Sign in to view Nathan’s full profile
or
New to LinkedIn? Join now
By clicking Continue to join or sign in, you agree to LinkedIn’s User Agreement, Privacy Policy, and Cookie Policy.
Denver Metropolitan Area
Sign in to view Nathan’s full profile
Nathan can introduce you to 10+ people at Amazon
or
New to LinkedIn? Join now
By clicking Continue to join or sign in, you agree to LinkedIn’s User Agreement, Privacy Policy, and Cookie Policy.
1K followers
500+ connections
Sign in to view Nathan’s full profile
or
New to LinkedIn? Join now
By clicking Continue to join or sign in, you agree to LinkedIn’s User Agreement, Privacy Policy, and Cookie Policy.
View mutual connections with Nathan
Nathan can introduce you to 10+ people at Amazon
or
New to LinkedIn? Join now
By clicking Continue to join or sign in, you agree to LinkedIn’s User Agreement, Privacy Policy, and Cookie Policy.
View mutual connections with Nathan
or
New to LinkedIn? Join now
By clicking Continue to join or sign in, you agree to LinkedIn’s User Agreement, Privacy Policy, and Cookie Policy.
Sign in to view Nathan’s full profile
or
New to LinkedIn? Join now
By clicking Continue to join or sign in, you agree to LinkedIn’s User Agreement, Privacy Policy, and Cookie Policy.
About
Welcome back
By clicking Continue to join or sign in, you agree to LinkedIn’s User Agreement, Privacy Policy, and Cookie Policy.
New to LinkedIn? Join now
Activity
1K followers
-
Nathan Huff shared thisThe cuffs are off. Yesterday the SEC killed the Pattern Day Trader rule. The $25,000 minimum — gone. The "four day trades in five days" tripwire — gone. If you have $500 in a brokerage account, you can now day-trade it into oblivion just as freely as someone with $25k. This is progress. It's also a brand new way to lose a lot of money. Not because day trading is evil — but most people lose money at this, especially starting out, especially without full-time focus. Which is exactly why I'm building Kantex: quantitative trading infrastructure for the people who just got handed the keys but never got the manual. The cuffs are off. Bring your own guardrails. https://lnkd.in/gAF2fmqS Fork this strategy to get started for free: https://lnkd.in/gb_A_66J #SEC #DayTrading #FinTech #Quant #StartupS&P 500 Institutional Accumulation Breakout by Sha Bang | KantexS&P 500 Institutional Accumulation Breakout by Sha Bang | Kantex
-
Nathan Huff shared thisA stale tag was quietly inflating my Cloud Run bill I use blue/green deploys on GCP Cloud Run. After every cutover my pipeline shifted 100% traffic to the new slot but never removed the tag from the old one. Cloud Run keeps tagged revisions addressable. That means instances stay warm even with zero traffic. Old revisions that should have scaled to zero were quietly consuming capacity around the clock. That was just the start. When I audited the rest I found: - Beta environment running without CPU throttling — paying for CPU even between requests - Orphaned revisions piling up across both projects, creeping toward the 1000-revision deploy limit Three fixes, one deploy: 1. Added --remove-tags to my CI/CD pipeline after cutover 2. Enabled CPU throttling on non-production environments 3. Wrote a cleanup script to prune untagged revisions weekly Cut my monthly Cloud Run spend by ~30%. I put together a audit checklist claude skill, that checks for things including zombie revisions, min-instance traps, the blue/green leak pattern and published it as a gist. Hope it saves someone a few bucks: https://lnkd.in/gd2VSkfTGCP Cloud Run Billing Audit — checklist for finding hidden costs from zombie revisions, stale tags, and blue/green deployment leaksGCP Cloud Run Billing Audit — checklist for finding hidden costs from zombie revisions, stale tags, and blue/green deployment leaks
-
Nathan Huff shared thisClaude Code’s new 'auto mode' is currently locked behind Team plans. I'm a solo founder. So I hacked together my own. Took about an hour. Here's what I learned. Anthropic published an engineering post explaining exactly how their classifier works. Turns out it's not magic — it's a 3-tier filter: - Read-only tools → always safe, instant approve - File edits inside your project → git has your back, instant approve - Shell commands → actually needs thinking https://lnkd.in/dVEKTu7Q For that last bucket, the real auto mode calls Sonnet 4.6. I call Haiku via claude -p (Claude Code's own CLI in non-interactive mode). No extra API keys, no new dependencies. The prompt is basically: "here's the working directory, here's the command — ALLOW or BLOCK?" Wrap that in a PreToolUse hook, add a regex fast-path for obvious stuff (rm -rf /, git push --force, base64 -d | bash), and you've got ~90% of auto mode for ~$0.12/month. The one real difference: Haiku like this doesn't see your conversation history, so it can't factor in what you explicitly asked for. It just looks at the command cold. In practice this hasn't mattered much, you just have to confirm even if you explicitly asked claude to run `rm -rf --no-preserve-root /`. Fail-closed: if the LLM times out or errors, it prompts instead of silently approving. 70 lines of bash. Works today. I guess like if this is helpful, but laugh if you are all in on `--dangerously-skip-permissions`Claude Code auto mode: a safer way to skip permissionsClaude Code auto mode: a safer way to skip permissions
-
Nathan Huff shared thisLaunching a product based on heavily leveraging Continuous AI (self-healing and improving user generated code) is a challenge that seemed laughable 4 years ago, in 2026 it is doable with mild hiccups such as: Overnight, a race condition in our message queues caused a massive compute spike. Vertex AI is still throttling us heavily, but we've successfully recovered. To keep systems running, we temporarily fell back to using Gemini Flash. Our next Continuous AI cycle will restore Pro for our most critical reasoning needs. This is turning into an interesting case study on Flash vs. Pro performance. I haven't benchmarked at this scale since moving to Gemini 3.1. While Continuous AI is vastly more complex than standard CI/CD pipelines, the magic of watching code progressively improve itself based on new inputs is truly incredible. When I first joined Amazon, I was captivated by the saying that "one-in-a-million edge cases happen daily at our scale." Less than a decade later, I never imagined I would be building systems that detect and fix those unanticipated incidents automatically. Site is back up, open beta is still free at https://kantex.ai/ Kantex.ai
-
Nathan Huff shared thisOne of my favorite teams in Amazon Devices (ABCD, Amazon Branded Connected Devices, the team that made the smart plug, air quality sensor, etc.) is hiring an Embedded Software Development Engineer in Denver. Ping me if you have questions. https://lnkd.in/gi-Qq5Yv
-
Nathan Huff shared thisI am proud to have worked on our latest tablet. It has a premium quality look and feel you won’t believe until you try it out in a consumer electronics retail store near you.Introducing the All-New Fire Max 11: Amazon’s Biggest and Most Powerful Tablet YetIntroducing the All-New Fire Max 11: Amazon’s Biggest and Most Powerful Tablet Yet
-
Nathan Huff posted thisMy team is hiring for pretty much every position from Manager, Quality Assurance Engineer, Software Engineer, Senior Software Engineer, and Senior Technical Program Manager. Been wanting to move to Denver but don't love the idea of packing all your stuff? Let me refer you to a recruiter to talk about our relocation packages! This is a link to the manager role: https://lnkd.in/eeVX476y
-
Nathan Huff liked thisNathan Huff liked thisIntroducing the Cadillac Championship Trophy! Standing 20 inches tall, the form captures motion, tension, and precision—hallmarks of both elite golf and Cadillac design.
-
Nathan Huff liked thisNathan Huff liked thisZscaler is proud to collaborate with Anthropic on Project Glasswing, supporting the responsible use of frontier AI to improve security outcomes for critical infrastructure. Claude Mythos Preview underscores the shift we are all facing. Frontier AI can accelerate vulnerability discovery and validation at machine speed, which means the old model of scan, patch, and hope will not scale. This is why I keep coming back to a core point. Zero Trust is not a feature. It is not a firewall with a new label. It is a fundamentally different architecture. Users never connect to the network and applications are never exposed to the internet. Endpoint context is evaluated and devices are verified before they connect. Data is protected the moment it is accessed. Every session, whether a human or an AI agent, is brokered one-to-one with verified identity in real time, with no lateral path to anything else. That architecture, delivered through the Zscaler Zero Trust Exchange, is how you reduce what attackers can see and reach in a machine-speed threat era. Read more: https://lnkd.in/geXHkK9A
-
Nathan Huff liked thisNathan Huff liked thisLooker got the most important thing right. A governed semantic model should sit between your data and your users. One source of truth. Business logic defined once. Every consumer -- dashboard, report, embedded app -- gets the same answers. That insight was correct. It's more correct now than it was in 2012. In 2012, the consumers were dashboards and reports. In 2026, the consumers are AI agents. And agents need meaning even more than humans do. An analyst can look at a column called "revenue" and know it's actually net revenue after returns. An agent can't. The model is what bridges that gap. Looker saw the future of data modeling before anyone else. What hasn't aged well is everything built around data modeling. LookML — a proprietary DSL locked inside a monolithic app. A dashboard-delivery paradigm that's being commoditized overnight. A product that cut off its modeling layer from the AI coding agent revolution. The insight was right. The packaging is what aged out. The lineage between LookML and Malloy is direct. The creator of LookML created Malloy -- open source, designed for the world Looker couldn't evolve into. Same core conviction about governed meaning. Different architecture for a different era. We have hired world-class data modeling and LookML experts to encode their expertise into skills -- modeling skills, analysis skills, migration skills. An AI agent with those skills can take your LookML, understand the intent behind it, and produce governed Malloy models in minutes -- not quarters. You keep everything you invested in the modeling. You lose the monolith that was holding it hostage. And now that model can go everywhere LookML couldn't. Into Cursor. Into Claude Code. Into MCP tools that any agent can consume. Into embedded analytics, Slack bots, and APIs -- all from one governed source. Open, portable, extensible. We're all on a journey right now. The technology is moving fast and nobody has the full map. But the teams that will move fastest are the ones that pack light -- leave your LookML baggage behind, carry the intent and the meaning forward, and stay open to where the road takes you. "Rain, Steam, and Speed -- The Great Western Railway" by J.M.W. Turner (1844)
-
Nathan Huff liked thisNathan Huff liked thisLots of “AI-generated BI dashboard” discussions lately. They remind me of a horse-drawn fire wagon hauling a shiny new steam engine. Real innovation bolted onto an old model. A fleeting moment in history -- and an interesting time to be alive.
-
Nathan Huff liked thisNathan Huff liked thisIt’s a note from DJ Patil, the first US Chief Data Scientist (2015-2017), summarizing his approach to tackling problems: Dream in years Plan in months Evaluate in weeks Ship daily --0-- Prototype for 1x Build for 10x Engineer for 100x --0-- What’s required to cut the timeline in 1/2? What needs to be done to double the impact?
-
Nathan Huff liked thisNathan Huff liked thisAt Amazon, being a jerk was rarely a career-limiting move. Unfortunately. When a VP was upset at a document he was reading, he told the Director (in front of their team) that they needed to send him a proper document in the next day or two. "Not this" he said, tossing it on the table, and walked out. I understand a high performance culture, and this document wasn't great. But he didn't have context on the team, or who wrote the document. He didn't wait to understand more, he just figuratively flicked off the team and walked out. He didn't and wouldn't get in trouble. Because he his teams delivered. A Director told me that too many people were asking to leave his team for mine because of his team culture issues (swear I'm not bragging, but it's a funny situation). He said I needed to block people from joining my team from his. Our VP agreed. The response was essentially, "This team is important. Your job is to convince them to stay." Not "Good job on having an attractive team", or "This Director is the problem." It's unfortunately a generally effective but inhuman incentive structure. When empathy doesn't appear in the performance management math, it gets treated as optional. It's not that being a jerk necessarily gets stuff done. But being too 'nice' can be seen as a weakness. And managers watch their leadership model behavior, and that moves through the company. There was a stark contrast when I later worked at Meta. People were regularly recognized in ratings discussions for helping colleagues succeed. At Amazon, if your personality was mentioned at all in a performance discussion, it was most likely a concern about a manager being described as "too nice." To learn more about what worked and what didn't inside Amazon, read on.4 (More) Good Things and 2 (More) Bad Things and 2 Misunderstandings About Working at Amazon: From 12+ Years in Management4 (More) Good Things and 2 (More) Bad Things and 2 Misunderstandings About Working at Amazon: From 12+ Years in Management
Experience & Education
-
Kantex.ai
******* * ********* ** *********
-
******
****** ******** *********** ********
-
*** ***********
******** ********
-
******** ****** ** *****
****** ** ******* ****** ******** ******* undefined
-
********* *******
****** undefined
View Nathan’s full experience
See their title, tenure and more.
Welcome back
By clicking Continue to join or sign in, you agree to LinkedIn’s User Agreement, Privacy Policy, and Cookie Policy.
New to LinkedIn? Join now
or
By clicking Continue to join or sign in, you agree to LinkedIn’s User Agreement, Privacy Policy, and Cookie Policy.
Languages
-
Russian
Elementary proficiency
-
English
Native or bilingual proficiency
View Nathan’s full profile
-
See who you know in common
-
Get introduced
-
Contact Nathan directly
Other similar profiles
-
David Janke
David Janke
Ocean Park Volunteer Rescue Squad
4K followersHampton Roads, Virginia Metropolitan Area
Explore more posts
-
Redis
293K followers
Learn how to build an image search app using a series of guided code commits with Developer Advocate Guy Royse. He’ll guide you through generating image embeddings using freely available Hugging Face models and storing them in Redis for fast, scalable semantic search. The workshop covers: ▶️ What embeddings are and how they work ▶️ How to generate image embeddings and store them in Redis ▶️ Building a semantic search app with Redis, Python, and JavaScript Get access here: https://lnkd.in/gbZ3imbz
44
-
Tensormesh
3K followers
We're excited to announce that Tensormesh has officially joined the Redis Partner Network. 🎉 Scaling AI in production isn't just a compute problem, it's a data coordination problem. Getting the right context to the right GPU at the right moment requires infrastructure that's fast, distributed, and resilient. That's exactly what this partnership is built to solve. Tensormesh optimizes LLM inference costs through intelligent KV cache management, keeping AI workloads running efficiently across distributed environments. By combining our inference orchestration with Redis's ultra-fast data layer, teams can now run AI systems that are responsive at scale, without the overhead. As our CEO Junchen Jiang puts it: "Scaling AI systems in production requires more than compute, it demands coordination, speed, and real-time data flow across distributed environments." We're proud to build alongside Redis as we push toward a future where production AI is both powerful and economically sustainable. 👉 Learn more about the Redis Partner Network: https://redis.io/partners/ #RedisPartnerNetwork #AIInfrastructure #LLMInference
59
3 Comments -
Tag1 Consulting, Inc.
959 followers
See Drupal differently, see LLMs differently. Part 3 of our Rust LLM series moves beyond abstractions and walks through a full GPT‑2 style model: from position embeddings and causal masking to GELU‑based MLPs and configurable model sizes (tiny to GPT‑2‑small). Engineers get both the big picture of transformer design and the low‑level details, tensor shapes, parameter growth, and real‑world forward‑pass timings that make experimentation practical. https://lnkd.in/gY4rCRNU #AIArchitecture #RustLang #Transformers #GenerativeAI #Tag1 #DigitalFirstResponders
-
Edge Newsletter
30 followers
AI jobs aren’t just about building models anymore. Companies need people who know: 📌 How to size hardware for workloads 📌 How to monitor a cluster’s health 📌 How to orchestrate jobs at scale 📌 How to run hybrid cloud + on-prem systems This certification proves you have those skills. 👉 https://lnkd.in/eUcvvpda #AIInfrastructure #JobSkills #TechCareers #tech
-
Edge Marketplace
7 followers
AI jobs aren’t just about building models anymore. Companies need people who know: 📌 How to size hardware for workloads 📌 How to monitor a cluster’s health 📌 How to orchestrate jobs at scale 📌 How to run hybrid cloud + on-prem systems This certification proves you have those skills. 👉 https://lnkd.in/eUcvvpda #AIInfrastructure #JobSkills #TechCareers #tech
-
Cypress AI
6 followers
AI jobs aren’t just about building models anymore. Companies need people who know: 📌 How to size hardware for workloads 📌 How to monitor a cluster’s health 📌 How to orchestrate jobs at scale 📌 How to run hybrid cloud + on-prem systems This certification proves you have those skills. 👉 https://lnkd.in/eUcvvpda #AIInfrastructure #JobSkills #TechCareers #tech
-
Krosspoint AI
2 followers
Kafka + Debezium: Real-Time CDC for Modern Government Systems Legacy batch jobs and nightly ETLs just don’t cut it anymore—especially in systems that demand real-time visibility. That’s where Apache Kafka + Debezium shine. Together, they enable change data capture (CDC)—capturing database changes as they happen and streaming them into downstream systems in real time. 🔄 Debezium listens to DB logs (Postgres, SQL Server, etc.) and emits structured change events 🚀 Kafka brokers those events reliably to services, dashboards, alerts, or data lakes 🛡️ Why this matters for federal systems: Enables live reporting, automated notifications, and rapid response Reduces load on OLTP systems—no heavy queries or joins Fits into zero-downtime modernization and microservices strategies Supports auditability and traceability for compliance We’ve used this pattern in public health and acquisition systems to create real-time integration between legacy data sources and modern platforms. #Kafka #Debezium #CDC #RealTimeData #Microservices #GovTech #KrossPoint #CloudNative #EventDrivenArchitecture #Modernization
2
-
Boden AI
31 followers
LLMs look impressive in demos — but become inconsistent once they're embedded in real workflows. BASE Studio helps teams fine-tune models from SFT to RLHF using data built for real-world use, not toy examples: 🔵 Multi-turn conversations that preserve context 🔵 Clear, followable instructions models can reliably execute 🔵 Code data that explains *why* decisions are made — not just the final answer 🔵 Reasoning data that strengthens step-by-step thinking For multimodal models, we also align prompts with visual inputs to reduce inconsistency in text-to-image and text-to-video generation. The result? Models that behave more consistently — even after deployment. Learn more 👉 https://boden.ai #EmbodiedAI #Robotics #AITraining #TrainingData #MachineLearning #DataPipeline #RobotLearning #MultimodalAI #Automation #ComputerVision #AIInfrastructure #Annotation
1
-
WITH Madrid
3K followers
We might be entering the era of overthinking AI 🤯. and it’s getting expensive 💸⚠️ Not because models aren’t powerful 🤖… but because we’re using them in ways that don’t justify their cost 💡❌ At WITH Madrid, we’re increasingly advising clients to challenge a common assumption: ➡️ More reasoning ≠ better outcomes ⚡ In many real-world use cases (classification 🗂️, routing 🛣️, simple decisions ✅), precision comes from design 🎨, not from internal monologues 🗣️💭 The implication is big 🚀: You don’t need a more expensive model 💰❌ You need a better architecture 🏗️
7
-
Domino Data Lab
56K followers
Tuning Spark, Ray, or Dask for ML isn’t one-size-fits-all. In a new Domino Blueprint, Principal Solution Architect Sameer Wadkar shows how API-driven scaling complements autoscaling to right-size clusters on demand—cutting idle spend and speeding bursts for GPU/I/O-heavy work. Practical steps, a lightweight Python client, and Helm install included. If you care about FinOps and MLOps, this is for you. Read the blueprint: https://lnkd.in/ga_qJkhC
7
-
Qdrant
57K followers
💡 Want Claude to give better technical answers using your own docs? In this 3-part series, Gergely Szerovay explains how to structure documentation for LLMs and build a complete RAG pipeline using Qdrant and MCP to feed that context into Claude Desktop. 📘 Start here with Part 1: https://lnkd.in/dfnrRdvE Next parts cover: – Chunking + embedding docs with LlamaIndex.ts – Setting up Qdrant for semantic search – Serving context to Claude via MCP – Multi-step and multi-query retrieval techniques for more accurate answers
18
-
AI Tech Suite
187 followers
Struggling with high LLM costs or slow AI responses? Redis just dropped some game-changing news that could revolutionize how we build intelligent agents! Redis is aggressively pivoting to become the foundational "real-time memory layer" for AI. At Redis Released 2025, they announced acquiring real-time data platform Decodable and launching LangCache, a new semantic caching service. These moves are designed to transform Redis into the ultimate engine for feeding fresh, contextual data to AI agents and ensuring they have persistent, scalable memory. This means massive efficiency gains for developers. LangCache alone promises to slash LLM API costs by up to 70% and accelerate response times by an incredible 15x by intelligently caching similar queries. Decodable will simplify complex real-time data pipelines, ensuring AI always has the latest context. Plus, deeper integrations with agent frameworks like AutoGen and LangGraph will make building robust, production-ready AI agents far easier. How do you see these updates impacting your AI development strategy or your industry? Share your thoughts! #AI #Redis #LLMs #RealTimeData #AIdevelopment Read more: https://lnkd.in/gQjBk3M8
-
Ververica
29K followers
Apache Fluss 0.8 is a pretty big deal. In this episode of 'Behind the Stream', Giannis Polyzos, Jark Wu and Leonard Xu break down why: → New storage tiering for cost optimization → Enhanced lakehouse integration (Iceberg, Paimon) → Performance improvements for high-throughput workloads → Production learnings from early adopters Jark and Leonard walk through real-world use cases where Fluss eliminated massive stateful operations and cut latency to single-digit milliseconds. If you're dealing with streaming storage at scale or trying to make your lakehouse actually real-time, this episode has answers. Listen now: 👉 Spotify: https://hubs.li/Q03Y6HPJ0 👉 Apple Podcasts: https://hubs.li/Q03Y6G-f0 #ApacheFluss #StreamingData #RealTimeAnalytics #FlinkForward
8
Explore top content on LinkedIn
Find curated posts and insights for relevant topics all in one place.
View top content