Elimnion Book a call

Faster co-folding pipelines, with the accuracy checked on your own targets.

Independent audits for ML and platform teams running Boltz-2 or OpenFold3. I find where the time and GPU cost go, then re-benchmark on your own targets to show the predictions held.

For ML and platform teams at drug-discovery biotechs and hosted co-folding platforms.

No NDA is needed for the first call. A short mutual NDA is available before you share anything specific.

your target complexes like the training data low high similarity to training set ligand RMSD < 2 Å success
The further your target is from the training set, the less a leaderboard tells you.

What I do

One audit, measured end to end on your own inputs.

Start here

Pipeline audit

I review the co-folding pipeline you already run, from inputs and preprocessing to model settings, filtering and compute use, and measure where time, memory and money go. Every speedup I recommend is checked against a fixed reference run.

  • Per-stage profile of time, peak GPU memory and cost per structure
  • Speedup validation report: RMSD, lDDT-PLI, ipTM deltas, PoseBusters pass rate and seed replicates
  • A prioritised list of changes, with expected effect and risk
  • The code, versions and configs to rerun everything yourselves

Benchmark on your targets

Public leaderboards rarely resemble your chemistry. I run the models on your targets and report pose accuracy, physical validity and where each model stops generalising, binned by similarity to the training data.

Performance engineering

After an audit, I can implement the agreed changes: batching, precision, memory layout, caching and preprocessing. Accuracy is re-benchmarked after each change, so a faster pipeline never quietly becomes a worse one.

Not sure an audit fits? Tell me where the pipeline hurts and I will say honestly whether it would help.

Book a 30-minute call

Why verification matters

Faster is easy to claim and hard to verify.

Time

Speedups are quoted for one stage, one dataset and one GPU. NVIDIA now gives away kernels, GPU search and inference containers, and the figures it quotes range from about 1.5x to 177x depending on stage, baseline and hardware. Its own caveats say results are specific to the dataset and hardware, and exclude steps such as preprocessing and retries.

Where it goes

In the AlphaFast preprint, MSA construction is over 95% of AlphaFold 3 wall time on both GPUs tested. Which stage dominates in your pipeline depends on your inputs and setup, which is why the profile comes first.

Scale

Memory limits decide which targets you can run at all. A public report describes an attention-kernel crash above about 2,000 residues (Boltz #452).

Accuracy

Faster settings are not automatically equivalent. A public report describes cached-MSA modes that degrade confidence separation on a peptide test set (Boltz #627). These are single reports, cited as examples of why checking matters.

Where AlphaFold 3 wall time goes L40S MSA construction 95.4% H200 MSA construction 95.8% 0% 100% gold = all other stages
Published figures from the AlphaFast preprint, not Elimnion measurements.

How results are verified

Numbers come with their hardware, versions and limits.

Against a reference

Each optimisation is compared with a fixed reference run plus seed replicates, on targets binned by similarity to the training set. Seed-to-seed variance alone is not treated as proof of equivalence.

Disclosed

Every figure states its GPU, software versions, baseline and what is excluded. Nothing is reported as a general speedup; it is measured on these inputs, against this reference, at this tolerance.

What it does not show

An audit checks computational equivalence and performance. It does not say whether a predicted pose is biologically true. That call stays with your scientists.

How an engagement works

Fixed scope, written up, and reproducible.

  1. STEP 1

    Scoping call

    Thirty minutes on the decision you need to make and the pipeline involved. No confidential detail is needed.

  2. STEP 2

    Mutual NDA

    If we proceed, a short mutual NDA is signed before you share anything specific.

  3. STEP 3

    Audit

    Scope and fee agreed in writing, then the audit. Uncertainty is reported, not hidden.

  4. STEP 4

    Report

    A written report, the reproducible code and configs, and a walkthrough call.

Ground rules

The answer is whatever the numbers say.

Independent

No affiliation with any model developer or hardware vendor, and nothing to sell you beyond the analysis.

Your data stays yours

By default the work runs in your environment, without uploading proprietary sequences or structures elsewhere. No client data is sent to third-party AI services without your approval, and everything is returned or deleted at the end.

Reproducible

Every result ships with code, versions and seeds, so your team can rerun and extend it without me.

About

Elimnion is an independent practice run by Pana, an ML engineer based in Brussels. Background in pharmaceutical sciences and drug discovery machine learning, with hands-on work on OpenFold and Boltz-2, and systems programming in Python, C and C++.

Engagements use only public data, my own tools and information supplied by the engaging client. I decline or disclose any conflict of interest before work starts.

Have a pipeline that is slow, costly or hard to trust?

Book a 30-minute call. Tell me where it hurts, and I will tell you honestly whether an audit would help.