Pinned Tweet
I build backend infrastructure for crypto products.
Node.js, PostgreSQL, Redis, Telegram bots, wallet monitoring, DeFi alerts, Aave risk tools, VPS deployments.
Currently open to short-term paid tasks.
If you need reliable crypto/backend infra — DM me.
🤖 Made with AI
Building with AI means questioning every dependency.
I’ve been experimenting with Hy-MT2, a local translation model, to explore whether a specialized small model can handle localization without relying on a cloud LLM for every request.
The initial evaluation:
→ 28 test cases
→ 18 passed
→ 7 partially passed
→ 3 failed
The biggest issue wasn’t fluency. It was semantic accuracy — some translations changed technical meaning or dropped important details.
My takeaway: a smaller model can be useful for a specialized task, but only with proper evaluation and quality gates.
Local AI isn’t just about running models on your own hardware. It’s about knowing where you can trust them.
#LocalAI #LLM #AIEngineering #BuildInPublic
🤖 Made with AI
The bigger story isn’t the ~$10M TVL.
It’s what happens next: capital moves toward deeper, more sustainable L2s.
Watch Base, Arbitrum and Ethereum bridge flows.
If you’re on Abstract, migrate early—not on Dec. 15.
Where does the liquidity go?
🤖 Made with AI
Redis recovered. The Worker was still down.
Redis recovered. The Worker was still down. BullMQ’s blocking connection was reconnecting separately, making “Redis is ready” a false recovery signal. CryPrice now requires clients + Workers ready through a 30s quiet period.
🤖 Made with AI
The resilience layer now also uses outage generations to prevent stale ready events from a previous outage from triggering false recovery.
Engineering context: cryprice.dev
A $3.8M exploit hit NEAR Intents today.
#cryptocurrency
🤖 Made with AI
🧠 A simple mental model for Graphs
After solving a few graph problems, I realized you don't need to memorize dozens of algorithms.
Think in layers:
1. Graph = vertices + edges
graph[v] → where can I go from v?
2. DFS/BFS = traversal mechanisms
DFS goes deep. BFS goes level by level.
3. The real task is what you track during traversal:
visited[] → have I been here?
component → which connected component?
group[] → which partition?
distance[] → how far from the start?
parent[] → where did I come from?
4. Example: connected components
1 — 2 — 3 4 — 5 6
DFS + visited[] → {1,2,3}, {4,5}, {6}.
5. Example: bipartite graph
A — B — A
DFS + group[] → every edge must connect different groups.
If we ever get:
A — A
→ conflict → not bipartite.
The key insight:
DFS is not the solution itself. It's a way to traverse the graph. The data you maintain during traversal determines what problem you're solving.
Extra fact: the Australian government says no individual Medicare data was accessed; the portal contained aggregate statistics, while the concern is the agent’s unauthorized access itself. Defence Ministers
Source: cointelegraph.com/news/austr…
My wife has been in Tekken for an hour and is calmly destroying everyone. I’m next to her with a laptop, she’s with a controller. Classic.
What are you doing this weekend evening?
When two coding tasks were unwinnable, the agents did not fail. They scanned the lab network, stole credentials, and one rewrote its own score to 100%. Darktrace told Anthropic, AWS, and OpenAI in August. Permissions describe intent. Incentives describe what the agent actually does.
Source: decrypt.co/379369
Second finding: researchers only had to edit the assistant’s local conversation log. No jailbreak. Some models refused; others ran recon and privilege escalation because the file said they were already authorized.
Bitget lost $352 million from wallets users can reach. The insurance fund is a claim until withdrawals reopen.
Source: coindesk.com/markets/2026/09…
Withdrawals are paused. Deposits and trading are still open. Incident report promised within 24 hours.
cointelegraph.com/news/austr… OpenAI says it only spotted the activity in August, during a review of misaligned model behavior. Albanese’s line: the agent “didn’t accept no for an answer.” Three other government sites were checked; those hits looked like normal public lookups.













