Superlinked

Open-source inference for agents.

[Superlinked @superlinked]

Most healthcare document AI can't be validated. The claims data is locked behind privacy walls, and synthetic data proves nothing.

CMS accidentally solved this: they published a documentation finding, complete with the answer. The rule required a face-to-face encounter within six months before delivery. The record showed seven. Result: insufficient documentation, payment recouped.

The regulator defined what "correct" means, so anyone can check the work.

Our example reproduces that finding end to end with five chained models, extracts the exact phrases ("L1851", "6 months", "7 months", "recoups payment"), and computes the one-month violation deterministically in code. All five required facts present, or the run halts rather than guesses.

The version of healthcare AI a compliance officer can sign off on: https://t.co/UlGMuceqdu

[Dayly @usedayly]

Welcome to Dayly, @superlinked. happy to be your GTM intelligence platform of choice.

more in 🧵

[Superlinked @superlinked]

SIE takes your JSON Schema and returns typed fields from a document using Qwen3.8 27B.

The hard part isn’t generating braces. It's making `injured_people: 1` mean the source actually says someone was injured.

Your parser won't catch a plausible lie.

[Superlinked @superlinked]

In SIE's extraction example, the source mentions a bruised passenger. Sonnet 5 returns `injured_people: 0`; Qwen3.8 27B on SIE returns `1`.

Both outputs look tidy. Only one belongs in the database.

[Superlinked @superlinked]

Qwen3Guard 4B on SIE returns a Safety verdict and harm categories. Your app decides whether to hold the request.

No dramatic refusal monologue required. Just a specialist doing one job before the main model speaks.

https://t.co/vUhEuwUvgD

[Superlinked @superlinked]

50,000 tokens per second only matters when the comparison conditions are clear.

We tested @topk_io `topk-embed-v1-small` through the Superlinked Inference Engine against its Hugging Face pipeline.

Both pipelines used one NVIDIA L40S, bf16 precision, identical inputs and warmed-up Python calls.

The results:

• 48,000 to 52,000 document tokens/s
• 1.5× higher document throughput
• 7.1× higher query throughput
• Single-query latency fell from 81 ms to 6 ms

Every test query still returned the same top document.

The useful lesson goes beyond this model. Model performance depends heavily on the runtime around it. Batching, GPU kernels, memory use and request scheduling can change the economics of an embedding workload without changing its checkpoint.

This matters for teams searching across text and page images, especially when late-interaction models need to run at production volume.

The article includes the benchmark setup, results by document length, quality checks and instructions for running both TopK models through SIE.

Read it here: https://t.co/uvx4tdKhHQ

#Embeddings #VectorSearch #AIInfrastructure #OpenSourceAI

[Superlinked @superlinked]

Love building with ML - and getting a room full of developers excited about it? We're hiring a DevRel Engineer at Superlinked to grow the community around our open source Inference Engine. You'll speak at conferences, run workshops and hackathons, build demos worth trying, and turn developer conversations into a better product. We're looking for someone who can ship code, tell a clear technical story, and bring people together, with a strong focus on the SF AI/ML scene.

Sound like you? Come join us!

https://t.co/YE2bQkBrZo

#hiring #devrel

[Superlinked @superlinked]

When a company restates its earnings, can your document AI tell the old number from the new one?

Get that wrong and you haven't automated analysis. You've automated a compliance incident.

The test case is real: Pathward Financial reported $45.096M net income, then restated it to $36.080M, with language that earlier filings "should no longer be relied upon." Three documents, one figure, and attribution is the entire product.

This example traces the number through all three SEC filings. Two extraction models must independently find every required span, or the run halts. Fail closed, by design: a pipeline that says "I couldn't verify this" is annoying; one that confidently reports a stale figure is dangerous.

Follow one number all the way down: https://t.co/J2W9QuxFMl

[Superlinked @superlinked]

Before your agent answers, SIE can run the request through Qwen3Guard 4B.

In one test, OpenAI Moderation returned `flagged: false` for a harmful request. SIE marked it Unsafe.

The bouncer works better at the door than in the incident report.

[Superlinked @superlinked]

You're laughing now. But come Monday, you'll be watching your coworker try to cram OCR, embeddings, a reranker and an LLM onto one GPU. All loaded at once.
Use SIE. Let the teddy bear keep his legs. 😉
https://t.co/0PzWE6W431

[Superlinked @superlinked]

Can AI read an insurance policy well enough to separate covered costs from excluded ones?

FEMA published a real appeal decision, so we could actually test it. The claim: $182,552 after a Lake Ontario flood. The barge quote alone: $181,832.94. What the policy covered: removing 12 to 15 cubic yards of stone from beneath the building. Not the barge, not the handling, not the disposal.

Four chained models do the reading (parse, extract, rerank, generate) and the output lands where FEMA landed, checkable line by line against the published decision.

It doesn't approve or deny anything. It turns a document pile into organized evidence so the adjuster starts from analysis instead of reading. That's the version of claims AI worth building right now.

Compare it against FEMA's decision yourself: https://t.co/HJDNVq9s5d

[Superlinked @superlinked]

In our 4,768 prompt safety eval, Qwen3Guard 4B on SIE reached 82.7% F1. GPT-6 Sol reached 76.1%.

Turns out the model checking the door doesn't have to be the biggest one in the building.

[Superlinked @superlinked]

Rekognition looked at a bed frame and confidently called it a coffee table.

SIE's SigLIP ranked "bed" first from the store's own labels.

100% confidence in the wrong furniture is still the wrong furniture.

[Abi Aryan @GoAbiAryan]

Wrapping up Cohort 2 and heading into Cohort 3 of the Inference Engineering and System Design course..

This is the kind of engineering community @LambdaAPI @modal @superlinked @baseten nurtures.

Thank u @TheZachMueller @charles_irl @philipkiely and Daniel for making it happen!

[Superlinked @superlinked]

On our 500 document extraction eval, Qwen3.8 27B on SIE got 83.5% of field values exactly right, versus 79.3% for Sonnet 5.

The schema can be valid while the answer is wrong. We measured the values.
Check this out:
https://t.co/dS1JTm2Zuq

[Superlinked @superlinked]

On 938 held-out product photos, SigLIP on SIE got type, color and material all right 72.3% of the time. Rekognition: 26.8%.

Estimated list-rate cost per million photos: $24 vs $1,750.

Finally, a labeling project finance won't hate. :)

https://t.co/YbgJyyptGu

[Tomasz Tunguz @ttunguz]

If a founder built a startup in the database market in the mid-2010s, 2% market share meant the company could IPO. No surprise the database was the most important software category in that era.

Inference is about to surpass it. https://t.co/D3pBcBh0Od

[Superlinked @superlinked]

Your catalog has its own types, colors and materials. SIE lets you encode the product photo and your candidate labels with SigLIP, then rank the matches.

No new training run because marketing decided "oatmeal" is now a color.

[Superlinked @superlinked]

SIE serves the same chat model through an OpenAI compatible API on Cloud or your own GPUs.

So your support agent can keep its integration when you change where it runs. The bot has enough identity crises without the infrastructure joining in.

https://t.co/NSYHhq1RBU

[Superlinked @superlinked]

We ran Qwen3.8 27B on SIE through 55 support conversations and 275 attempts to bend the rules. It followed every tested policy.

Customers can be persuasive. Your system prompt shouldn't be.