Niels Bantilan

Building @union_ai, creator of pandera for data validation

[Niels Bantilan @niels_ai]

Ran a comprehensive benchmark using Jev in a System 1 + System 2 pipeline on @union_ai

🧐 System 1: raw data parsing, classification, scoring, noul-ing
🤔 System 2: reasoning, planning, text generation

Three pipelines:
- Legal contract draft review
- Code review
- Customer support ticket handling

Latency with Jev: 132.4s (528.1s without) - x4 speed-up
Cost / run with Jev: $0.00124 ($0.01185 without) - ~10x cost reduction
Decision stability with Jev: 100% (88% without)

Code (a lot more lines, because it's a comprehensive benchmark 😅)

https://t.co/pFa5RfkPoO

[Niels Bantilan @niels_ai]

Yep! `hlock pin --user` snapshots your user-level setup into harness.lock: harness version, models, skills, instructions, config, and every MCP server (command/args or URL, env, scope). (by default `hlock pin` only grabs project-scoped configuration). Secrets become placeholders and never go into the file.

[Niels Bantilan @niels_ai]

1/7: I got a new laptop. Which MCP servers did I even have?

It took a while to get settled into a nice Claude Code setup: MCP servers, skills, a CLAUDE.md, plugins, and my preferred default model. None of it was written down anywhere.

So I built hlock, which works like uv for your agent setup. 🧵

[Niels Bantilan @niels_ai]

7/7: try it out

hlock is Apache-2.0 with a Rust core 🦀. It works with #ClaudeCode, #Codex, #Cursor, #Pi, #OpenCode and more.

pip install hlock

https://t.co/bZwsUC1z69

Pin your setup, commit it, and reply with the link. We'll feature one lock every week.

[Niels Bantilan @niels_ai]

6/7: beyond pinning

A lock is also a starting point:

⇄ `hlock migrate --to codex` moves your Claude Code setup to Codex
🏠 `hlock localize` runs the same setup on local models 📈 hlock resolve and hlock tune measure setups on your own tasks and lock the winner

[Niels Bantilan @niels_ai]

7/ Read more in the blog https://t.co/KXbdlXIZ29

[Niels Bantilan @niels_ai]

1/ The Flyte devbox can now use your Mac's GPU.

Start it with 'flyte start devbox --apple-silicon'

Then request resources with 'flyte.Resources(gpu="AppleSilicon")'

On my M5, I get 13.4 TFLOPS on "mps" vs 1.6 on the CPU, measured from inside the task.

[Niels Bantilan @niels_ai]

6/ There's a proof task that fails unless macOS itself reports the GPU busy during the work (IORegistry, the counter Activity Monitor draws).

[Niels Bantilan @niels_ai]

Read more in the full blog: https://t.co/b2WljiEbqp

Docs:
- https://t.co/1O3y7SfYR1
- https://t.co/OYBMqTy4tE

[Niels Bantilan @niels_ai]

💥 A 9-hour AI pipeline fails at hour 8. Whether you pay for the first 7 hours twice comes down to which verb you reach for next.

Flyte 2 has three, and they differ on whose code runs and what re-executes.

[Niels Bantilan @niels_ai]

All three need a remote cluster, because they work from the platform's record of the source run.

[Niels Bantilan @niels_ai]

The catch is that only the task function's own body is hashed.

If you edit a helper it calls or a module-level constant it reads, the task gets reused with its old result. --force-rerun-action is how you make it run.