Writing
Notes on technology, applied engineering, product development, academics, experience, and literature.
No posts in this category yet.
2026 2
-
Designing a safe AI companion: having memory while remaining controllable
Jun 23 · 5 min read
An AI companion with memory doesn't just mean saving as much as possible. For safety, we need to clearly design what memory is saved, who can view it, when it's deleted, and when humans must be brought into the control loop.
-
Snow AI Companion: what to consider when designing child-safe AI?
Jun 22 · 6 min read
A design note for Snow AI Companion: if AI targets children, especially those needing support in communication and learning, the product must start with safety, control, and gentle companionship.
2025 3
-
AI Agents shouldn't start with agents, but with workflows
May 8 · 6 min read
Before talking about AI Agents, tool use, or autonomous systems, we should start with a real workflow: who does what, where the data is, what steps need automation, and what steps need human control.
-
Why AI Engineers need to understand business workflows, not just models
Apr 17 · 6 min read
An AI Engineer doesn't just need to know prompts, RAG, or models. To create real value, they must understand the workflow, data, users, and how the system operates in practice.
-
The process of reviewing a new open-weight model like an AI product engineer
Feb 6 · 7 min read
It's not just about looking at benchmarks. Here is a practical checklist for reading an open-weight model: model card, license, architecture, context, inference, evaluation, and product fit.
2024 5
-
Reading new LLM papers: which trends matter for AI Engineers?
Nov 21 · 7 min read
A pragmatic approach to reading LLM papers: ignore the hype, and look for trends that actually affect how we build AI products.
-
Optimizing LLM reasoning is not just about training bigger models
Oct 8 · 6 min read
How inference-time scaling, self-consistency, verifiers, and reasoning budgets improve LLM quality, and when not to use them due to high costs.
-
Why a good model doesn't necessarily create business value?
Apr 11 · 5 min read
A note from the Booking.com case: model performance and business performance are two different things, especially when ML goes into a real product.
-
Operate what you build: lessons for small AI products
Mar 14 · 4 min read
Analyzing the ownership mindset in production engineering and how to apply it to AI workflows, CRMs, dashboards, and small products.
-
OnePiece at Shopee: How LLM-style reasoning enters a ranking system
Jan 18 · 3 min read
Notes on the OnePiece paper and how context engineering and reasoning are introduced into an industrial cascade ranking system.
2023 13
-
Designing a Remote Terminal/WebSocket Dashboard for AI Agents
Dec 7 · 7 min read
Technical notes on designing a real-time dashboard to monitor and control AI agents running in the terminal via WebSocket.
-
How to Migrate a Test/Evaluation Pipeline for AI Workflows Without Breaking Production
Nov 16 · 6 min read
Practical notes on how to change prompts, models, or evaluation pipelines for AI workflows while keeping the system stable.
-
Evaluating chatbot/AI workflow: benchmark, verifier, LLM judge and real-world test cases
Oct 5 · 7 min read
A practical note on LLM evaluation: benchmarks are just the starting point, while a real product needs test cases, verifiers, human review and metrics tied to the workflow.
-
API Design for CRM: Leads, Contacts, Deals, Quotes
Sep 14 · 5 min read
Notes on designing a REST API for a small CRM, focusing on leads, contacts, deals, quotes, and real-world sales workflows.
-
When Does It Make Sense to Use Redis Cache for a CRM Dashboard?
Aug 3 · 6 min read
Practical notes on using Redis cache to optimize a CRM dashboard: when to use it, what to cache, invalidation strategies, and how to avoid common pitfalls.
-
RecSysOps: operating a recommender system after deployment
Jul 13 · 5 min read
Notes from Netflix RecSysOps on operating recommender systems: issue detection, issue prediction, diagnosis, and resolution when a recommendation system goes into production.
-
Monitoring AI workflows: learning from Netflix but applying at small scale
Jun 8 · 3 min read
Designing just enough monitoring for small AI workflows: logs, metrics, traces, evaluation, and business signals.
-
From Discord to CRM: how to store activity logs and conversation history?
May 11 · 5 min read
Drawing lessons from how Discord handles message storage to design activity logs and conversation history for a small CRM.
-
When the database starts exhausting the on-call team
Apr 18 · 3 min read
From Discord's case of storing trillions of messages, the post looks at the signs that a database is no longer suited for the workload.
-
Building a CI/CD Pipeline for Small AI/Computer Vision Projects
Mar 23 · 6 min read
A practical pipeline for a small AI or Computer Vision project: test, build Docker image, run basic eval, and deploy more securely.
-
Vision, object detection and segmentation: understand correctly before putting into production
Mar 2 · 4 min read
Short notes on object detection, segmentation, and why real computer vision problems often need more than a bounding box.
-
From Mask R-CNN to Mask R-CNN2Go: when computer vision research goes into production
Feb 9 · 3 min read
A note on how a computer vision idea from research can be optimized to run on real devices.
-
Computer vision on edge devices: learning from Meta Mask R-CNN2Go
Jan 17 · 3 min read
A short post on edge AI, latency, and the questions to answer before bringing computer vision down to real devices.
2022 14
-
Microservice Architecture Patterns for Scalable ML Systems
Dec 1 · 6 min read
Practical notes on how to break a Machine Learning system into smaller services to make it easier to deploy, monitor, and scale.
-
Feed ranking and feedback loops: lessons for product builders
Nov 3 · 4 min read
Explaining the feedback loop in recommendation systems and how product builders should design signals, metrics, and guardrails.
-
Tesla camera-only perception vs sensor fusion: what is the real trade-off?
Oct 11 · 3 min read
Careful notes on camera-only perception, occupancy perception, and sensor fusion in autonomous driving.
-
Does TikTok use Kafka/Flink for real-time recommendation?
Sep 6 · 3 min read
A fact-checking post: what public sources allow us to say and not say about Kafka/Flink in TikTok's recommendation.
-
TikTok-style architecture: should be understood as a feed ranking pattern, not an internal diagram
Aug 16 · 3 min read
A cautious post about TikTok-style architecture: using only public sources to derive product patterns, without asserting internal architecture.
-
TikTok-style recommendation: why is architecture as important as the algorithm?
Jul 28 · 4 min read
Analyzing TikTok-style feeds through official sources and academic research to understand the role of platform architecture in recommendations.
-
LightSAGE at Shopee: GNN for item retrieval in e-commerce ads
Jul 7 · 3 min read
Analyzing the LightSAGE paper on how Shopee uses graph neural networks for item retrieval in recommendation ads.
-
Shopee and distributed tracing: how to watch a request go through microservices?
Jun 14 · 3 min read
Analyzing Shopee's use of ClickHouse for distributed tracing and lessons for multi-service systems.
-
When should you leave MongoDB/Cassandra/SQL? Lessons from Discord
May 26 · 4 min read
Analyzing the database migration decision through the Discord case and how to choose a database based on access patterns, operational costs, and trade-offs.
-
How does Discord store trillions of messages?
May 3 · 6 min read
Reading the Discord case of migrating from Cassandra to ScyllaDB and drawing lessons on databases, hot partitions, migration, and latency.
-
How Netflix uses ML to optimize streaming quality?
Apr 12 · 5 min read
Notes from the Netflix Tech Blog on how Machine Learning is used to predict streaming quality, reduce playback errors, and improve viewer experience.
-
Netflix Microservices: why is observability more important than you think?
Mar 24 · 4 min read
Reading Netflix's post on microservices to understand why distributed systems need multi-level observation: request flows, bottlenecks, and instance-level metrics.
-
Designing a ranking system: from Airbnb to CRM search
Mar 8 · 4 min read
Drawing lessons from Airbnb Search and Embedding-Based Retrieval to design a CRM search/ranking system in a practical way.
-
What blocks does an ML platform need? Learning from Uber Michelangelo
Feb 17 · 3 min read
Analyzing Uber Michelangelo to understand that a production ML platform needs data, training, deployment, prediction, and monitoring.