Hi, I'm
Nguyễn Lý Thành
Backend Engineer
Backend engineer, 6+ years. Design doc to production — I own the technical side end-to-end.
Systems I've Built
Built to shipExperience
- 05/2022 – PresentHCMCBackend DeveloperParcel Perform CO. LTD
- Notification & Auth — built and scaled a multi-channel notification pipeline (SMS, Email, Webhook) to 20M+ records/day; owned platform auth across accounts, roles, permissions, and org structures; led the Payment Gateway re-architecture end-to-end with DHL and Reach Bank integrations; authored 5+ technical designs.
- Parcel Monitor Tracking & Email Sync — re-architected Email Sync to 10M+ emails/day; built Gmail sync pulling shipment data into PMT via rule-based and AI email extraction; maintained User Management and Email Sync services; cut Firebase cost/storage and lifted test coverage past 90% across 3 technical designs.
- AI Platform — Tech Lead for Campaign Engine, AI Ecommerce Visibility, and Sales Agent on AWS Bedrock; built MCP lead-gen and ELK monitoring.
- 03/2021 – 07/2022HCMCBackend DeveloperJobHopin CO. LTD
- Led monolith → microservices migration for the main recruitment platform.
- Built JOB-AI LYSIS solo (Django, DRF, Celery, AWS, MySQL, Docker) — published on SAP Store via SAP contract.
- BDD automation framework in Go/Godog covering 100+ scenarios.
- 02/2019 – 02/2021HCMCFull-Stack DeveloperKyanon Digital CO. LTD
- Front-end for AI-powered building facade inspection (30+ defect types) with ReactJS + Sketchfab 3D.
- AML real-time transaction screening (Django, Celery, AWS).
- 4-service identity verification: Facecompare, OCR, Digital Signature, Auth.
Featured Project
TSOnline TNT
LiveSolo-built and still running a real-time game server — full-stack, full ownership.
Key Achievements
High-Scale Systems
Designed and operated pipelines processing 20M+ records and 10M+ emails per day with stable throughput.
Parcel Perform Footprint
~2,200 commits across 45 repos, with deep ownership of data pipelines, GenAI prospect platform, analytics, file ingestion, integration services, and streaming core.
AI Products
Tech Lead on Campaign Engine, AI Visibility, and Sales Agent — LangChain/LangGraph/LangSmith on AWS Bedrock with Guardrails.
Performance Wins
Drove PMT shipment search latency from 60s down to 23s.
Full Ownership
TSOnline TNT — sole backend + ops, in production for 6+ years.
Blog
Notes on backend, AI agents, and shipping things that hold up in production.
Together AI Just Raised $800M at an $8.3B Valuation — Aramco Ventures Is Betting Sovereign Capital on Open-Source AI Infrastructure
On July 1, 2026, Together AI closed an $800 million Series C at an $8.3 billion valuation — more than 2.5x what it was worth 16 months ago — led by Aramco Ventures, the venture arm of Saudi Aramco, with Nvidia, Vista Equity Partners, and General Catalyst also participating. The company says annual bookings now exceed $1.15 billion and that enterprises are cutting inference costs up to 60x versus closed-model APIs by running open-weight models like Llama, DeepSeek, and Qwen on its platform. Here is what the round actually signals about where AI infrastructure value is concentrating, and what to check before your next model-hosting decision.
GitHub Copilot CLI Went Multi-Model in One Week — And Nobody Noticed the Chinese Coding Model It Just Added
In three GitHub Copilot CLI releases between June 30 and July 1, 2026, GitHub added Claude Opus 4.8 Fast, Claude Sonnet 5, and Moonshot AI's open-weight Kimi K2.7-Code — a roughly 1-trillion-parameter model that beats Opus 4.8 on MCP Mark Verified while burning about 30% fewer reasoning tokens. Each changelog entry is a single line. The implications for how enterprise engineering teams vet, budget for, and secure their AI tooling are not. Here is what actually changed, and what to check before your next sprint planning.
Claude Fable 5 Returns After 19 Days Offline — But Its New Safety Classifier Will Flag Your Routine Code Too
On July 1, 2026, Anthropic restored global access to Claude Fable 5, ending a 19-day suspension triggered by a US export control order after Amazon researchers found the model would produce exploit code once a security-review request was rephrased as "fix this code." The fix is a new cybersecurity classifier that blocks the technique in over 99% of cases — but flags routine coding and debugging more often too, silently falling back to Opus 4.8 when it fires. More than 100 security experts, led by Katie Moussouris, dispute this was ever a real "jailbreak." Here is what changed, what's still disputed, and what to check before July 7.