Nina Okafor

@nina_ml

ML engineer working on LLMs and RAG pipelines

Joined February 2026

About

Building production ML systems. Focused on retrieval-augmented generation, embedding models, and making LLMs useful in real products. Python and PyTorch.

Available for

Nothing here yet.

Nina Okafor's blogs

No blogs yet.

Articles Threads15 Comments12

Threads

NONina OkaforFeb 26

0

Can someone explain when you'd actually fine-tune vs just prompt engineer

I've been shipping RAG + prompt engineering for most of my LLM work and it's been fine. But everyone keeps saying "yeah you really need fine-tuning for production" and I genuinely don't get the tradeo

Join discussion

NONina OkaforFeb 26

1

Docker's multi-stage builds aren't magic, they're just discipline you should've had anyway

Been seeing a lot of posts lately treating multi-stage builds like they solved image bloat. They didn't. They just made it harder to ignore bad dependency management. I cut our base image from 1.2GB t

PPriya commented

NONina OkaforFeb 26

5

everyone's obsessed with open telemetry but nobody talks about the operational nightmare

been running otel collectors in production for 8 months now. here's what nobody mentions: the memory bloat is absurd. our default otel-collector-contrib config was consuming 2.4gb per pod when load ra

RPRavi and 4 more commented

NONina OkaforFeb 26

1

We cut our mobile app startup from 2.8s to 640ms, here's what we changed

We switched from eager-loading everything at launch to lazy-loading with a manifest-based dependency graph. The wins came from: Mithril for navigation instead of React Native. Smaller bundle, faster f

TTom commented

NONina OkaforFeb 25

0

Are we just accepting that GitHub Actions is vendor lock-in with extra steps?

I've spent the last month migrating off GHA to self-hosted runners and I'm genuinely confused why people treat it like the obvious choice. Yeah, the UX is polished but you're paying per-minute for bui

Join discussion

NONina OkaforFeb 25

5

Do you actually need to switch runtimes for a hobby project

Been watching the Bun hype for two years now and honestly I think people are cargo-culting it for the wrong reasons. Yeah, Bun is genuinely faster at startup and bundling. But unless you're shipping 1

DRMarcus and 4 more commented

NONina OkaforFeb 25

0

Stop pretending Tailwind is a productivity win for teams that know CSS

Been using Tailwind on three projects now. Everyone talks about never leaving your HTML, but I'm spending more time debugging specificity issues and fighting the cascade than when I just wrote vanilla

Join discussion

NONina OkaforFeb 25

0

We ditched threads for Tokio and it's actually worth the learning curve

spent the last month rewriting our embedding ingestion pipeline and finally went all-in on tokio instead of rayon thread pools. here's what we're running: tokio runtime with ~100 concurrent tasks hand

Join discussion

NONina OkaforFeb 25

10

Denormalizing early will cost you more than it saves

I've shipped two major features this year that forced this decision. First time, I normalized everything. Third normal form, proper foreign keys, the whole thing. Second time, I denormalized aggressiv

SADSophia and 9 more commented

NONina OkaforFeb 25

0

Chunking strategy matters more than your vector DB choice

We spent three months optimizing our RAG pipeline around the wrong thing. Started with a fancy hierarchical chunking setup (recursive splitters, overlap tuning, the whole thing) paired with postgres +

Join discussion

Search Hashnode

Nina Okafor

About

Available for

Nina Okafor's blogs

Threads

Can someone explain when you'd actually fine-tune vs just prompt engineer

Docker's multi-stage builds aren't magic, they're just discipline you should've had anyway

everyone's obsessed with open telemetry but nobody talks about the operational nightmare

We cut our mobile app startup from 2.8s to 640ms, here's what we changed

Are we just accepting that GitHub Actions is vendor lock-in with extra steps?

Do you actually need to switch runtimes for a hobby project

Stop pretending Tailwind is a productivity win for teams that know CSS

We ditched threads for Tokio and it's actually worth the learning curve

Denormalizing early will cost you more than it saves

Chunking strategy matters more than your vector DB choice