Anil Kurmi

🚀 AI Architect
⚙️ Distributed Systems & Microservices
🤖 Writing a book on AI Agent Architecture
✍️ Creator: PracticeOverflow
Let's build scalable tech.

[Anil Kurmi @practice108om]

https://t.co/b6QqYIB1MP

#ai #layoffs

[Anil Kurmi @practice108om]

https://t.co/MqV33wkeYJ

#ai #ailayoff #layoffs

[Anil Kurmi @practice108om]

Your p50 dashboard is lying about your agents.

A 20-step agent hits tail latency ~18% of the time. Not bad code. Just math: ~1% odds per call, compounded over 20 steps.

The median never sees it. Users do.

#SoftwareArchitecture #DistributedSystems #SRE #Reliability

[Anil Kurmi @practice108om]

Apple shipped the new Siri this week. The brain inside it belongs to Google — a 1.2T-param Gemini model it rents for ~$1B/yr.

The most integrated company on earth chose to rent intelligence, not build it.

#AI #LLM #Apple #Google #GenAI

[Anil Kurmi @practice108om]

AI won't take your job. It's quietly changing what your work is worth.

I spent months on how people actually survive automation waves. Wrote it up as a free 6-part series — no paywall. Thread of all six 👇

#AI #FutureOfWork #CareerGrowth #Upskilling

[Anil Kurmi @practice108om]

Part 5 — The Builder's Playbook. 30/90/365-day moves for engineers, freelancers & IT pros:
https://t.co/zYdNEf1vnC

Part 6 — The Knowledge Worker's Playbook:
https://t.co/R1RnJwJltp

Which of your tasks are gaining value as AI improves — and which are quietly losing it?

[Anil Kurmi @practice108om]

AI won't take your job. It's quietly changing what your work is worth.

I spent months on how people actually survive automation waves. Wrote it up as a free 6-part series — no paywall. Thread of all six 👇

#AI #FutureOfWork #CareerGrowth #Upskilling

[Anil Kurmi @practice108om]

Part 3 — The Structural Shift. What the labor data really says:
https://t.co/xDHt3j12WQ

Part 4 — Global & Industry Patterns. The survival map across 14 sectors:
https://t.co/uocQTao8yI

[Anil Kurmi @practice108om]

"Will AI take my job?" is the wrong question.

Ask: which of my tasks got cheaper, which got more valuable?

AI reprices tasks inside a job, not the whole job. Move toward work AI can't sign off on 👇

https://t.co/al2Dk5UCdf

#AI #FutureOfWork #AICareers #TechCareers

[Anil Kurmi @practice108om]

AI job loss is not one story.

Europe: hiring cuts.
China: robot supply chain.
India: mid-tier IT beats scale.
Philippines: BPO pivots to capability.

Same AI. Different labor markets.

The moat is capability per person.

#AI #FutureOfWork #Automation #IndiaIT #TechCareers

[Anil Kurmi @practice108om]

The best systems don't hide risk.
They move it where teams can control it.

This week: containment over fake fixes, graceful overload over heroics, agent runtimes over tool sprawl, living runbooks over PDFs.

#SoftwareArchitecture #DistributedSystems #SRE #AIInfra

[Anil Kurmi @practice108om]

AI Weekly W23: Microsoft shipped 7 in-house MAI models. Uber capped AI coding at $1,500/mo. Ladybird closed PRs. Hackers asked Meta AI for Instagram accounts and got them.

Capability is racing. Trust, budget, and auth aren't.

#AIWeekly #AI #Microsoft

[Anil Kurmi @practice108om]

@coolcoder56 Curious — was the system design round LLD or HLD heavy? And did they push on failure modes (worker death, partial writes) or stay on the happy path?

[Anil Kurmi @practice108om]

Job-scheduler LLD: 40 min burned on Redis vs SQS. Real rubric: what happens when a worker dies mid-task? Lease + heartbeat + idempotent retry + DLQ = yes. Miss it, stay senior. What failure mode did YOUR interviewer drill? #SystemDesign #InterviewPrep #BackendDev #Career

[Anil Kurmi @practice108om]

@RaulJuncoV Yes — and the catch is even teams with evals get burned by stale golden sets. Switched to property-based fuzzing on tool-call JSON last week, caught 3 schema drifts our 200-row set was happily green on.

[Anil Kurmi @practice108om]

Killed our 200-row golden set for property-based evals on tool-call JSON. Hypothesis fuzzed 5K variants, caught 3 schema drifts the GPT-4o→GPT-5 snapshot tests missed. $0.40/run. Anyone else ditching golden sets for fuzzing? #LLMOps #Evals #AI #SoftwareEngineering

[Anil Kurmi @practice108om]

@e_opore Most "fault-tolerant" systems fail the cross-region test. We thought sync multi-region replication was tolerance — turned out to be 280ms on every write. Async + per-region idempotency keys is what actually survives a region going dark.

[Anil Kurmi @practice108om]

Cross-region sync replication: 280ms per write. Switched to LOCAL_QUORUM + async cross-DC with per-region idempotency keys — same correctness, 18ms p99. Strong-everywhere is almost never what you actually need. #DistributedSystems #SystemDesign #BackendDev #Cassandra

[Anil Kurmi @practice108om]

@DbrxMosaicAI Curious how you scoped the fix — schema-typed tool outputs, post-hoc summarization, or both? Wrapping returns in Pydantic before the planner sees them got us a 27× shrink, but summarization felt more honest.