Alakh Krishnani

AI builder | writer | reader | Liverpool FC fan | movie enthusiast

[Duolingo @duolingo]

they named a phone after me and made it bend over 😭

[Chad Infi𓄿 @chad_infi]

Our BrahMos hit the target again.

[Alakh Krishnani @holyalakh]

@puremaureen @grok what is the name of this show / movie?

[Alakh Krishnani @holyalakh]

Tried Google's TurboQuant KV cache compression on an i5 + 20GB RAM, no GPU. It runs. But the latency tradeoff on CPU-only is brutal. Compression saves memory, costs you time. Worth it if you're RAM-bound, not if you're latency-sensitive.

[Alakh Krishnani @holyalakh]

Root keys on a solo project feel faster. Until something goes wrong. Scoped IAM user + named profile. Takes 15 minutes to set up.

Saves you from the worst day of your career.

[Alakh Krishnani @holyalakh]

I see too many solo devs running AWS with root keys.
Here's exactly how I set up a scoped IAM user for one AI project. Minimal permissions, easy to rotate.

[Alakh Krishnani @holyalakh]

Why this matters: if the keys leak, the blast radius is one S3 bucket and one EC2 instance type in one region.

Rotating takes 2 minutes. Revoking takes 10 seconds.

[Alakh Krishnani @holyalakh]

Every AI procurement demo uses clean sample data. Real spend data has duplicate invoices, inconsistent supplier names, missing GL codes.

The AI isn't the hard part. Data normalisation is. That's where implementations actually die.

[Alakh Krishnani @holyalakh]

Most posts comparing AWS GPU instances focus on VRAM. For LLM inference, g4dn beats p3 on cost-per-token by a wide margin. p3 has more VRAM but you're paying for training capacity you don't need. The decision tree is simpler than people make it.

[Alakh Krishnani @holyalakh]

First question enterprise clients ask when you demo an AI tool isn't about accuracy.
It's "where does my data go?" Design the audit trail before the UI.

[Alakh Krishnani @holyalakh]

Tried Kiro on Ubuntu. The spec-driven workflow is a genuine shift from VS Code + Copilot. You define what you want built before writing code. Whether that's better depends on how structured your project already is.

[Alakh Krishnani @holyalakh]

Every "one afternoon" automation is a debugging marathon through undocumented edge cases. I estimated 1 hour. It took 5. Shipping the honest version matters more than the clean version.

[Alakh Krishnani @holyalakh]

built a daily AI content engine on a $200 NAS. Drafts for 5 platforms every morning. ~$0.02/day. It took 6 failures before a single draft landed. Full breakdown 🧵

[Alakh Krishnani @holyalakh]

What's actually running now: NAS cron → curl to n8n webhook → Anthropic API → Notion. Platform-specific formatting for 5 channels. 5 min review each morning. ~$0.02/day.