Experiential Labs turns user traffic into model training data

Experiential Labs has launched an open-source AI gateway that turns production traffic into training data. The pitch, posted to Product Hunt, is compact: “Open source AI gateway turning traffic into a better model.” The mechanism is the interesting part. Most gateways sit between an application and a model provider, handling routing, caching, rate limits, and cost controls. Experiential Labs adds a feedback loop on top of that plumbing. Every request that flows through the gateway becomes a candidate for the next fine-tuning run.

The idea is not new in isolation. The AI industry has spent two years chasing the data flywheel, the loop where usage generates the examples that improve the next model. OpenAI, Anthropic, and Google all run variants of this internally. What Experiential Labs is doing is different in one specific way: it is packaging the flywheel as an open-source tool that any team can run on its own traffic. That changes who gets to participate in the loop. It also raises questions about what happens to the data that flows through it.

What the gateway actually does

The Product Hunt listing is thin on technical detail, which is typical for a launch page. What is clear is the core architecture. The gateway intercepts requests headed to a model provider, records the inputs and outputs, and feeds those pairs back into a training pipeline. The stated goal is a model that improves continuously as it serves real users. This is the difference between offline evaluation and online learning. A model that only learns from curated benchmarks improves at benchmarks. A model that learns from live traffic improves at the tasks people actually ask it to do.

The open-source framing matters here. A closed gateway that does this would be a vendor lock-in play. A company would route all its traffic through the tool, the vendor would accumulate the data, and the customer would be stuck. The open-source version inverts that relationship. The team running the gateway owns the data, owns the training pipeline, and owns the resulting model. The tool is the commodity. The data is the asset.

That is a meaningful shift in the economics of AI infrastructure. The past two years have been defined by the concentration of training data in a handful of labs. Experiential Labs is betting that the next wave of model improvement comes from the long tail of production traffic that currently evaporates after each inference call. Every request that a customer makes to a model is a piece of signal. Most of that signal is currently discarded. The gateway is a way to capture it.

The data quality question

The obvious objection is that production traffic is messy. Users ask vague questions. They make typos. They submit malformed inputs that no reasonable training set would include. The counterargument is that this messiness is exactly the point. A model that only trains on clean, curated data learns to expect clean, curated inputs. Real users do not behave that way. The gateway captures the full distribution of real usage, including the edge cases that benchmarks never cover.

There is a second objection that is harder to dismiss. Production traffic is not just messy, it is also private. Every request that flows through a gateway carries whatever the user was working on. A legal team asking a model to draft a clause is sending confidential information through the pipeline. A healthcare startup asking about patient symptoms is sending protected health information. The gateway records all of it by default. The training loop only works if the data is retained, and retention is precisely what most privacy policies forbid.

Experiential Labs is not the first company to face this tension. OpenAI and Anthropic both offer API customers the option to opt out of training on their data. The default for enterprise customers is no-training. The gateway inverts that default. Every request is training data unless the operator configures it otherwise. That is a product decision with real consequences for anyone who deploys it without reading the fine print.

Who this is for

The target customer is not a frontier lab. The target customer is a mid-size company that runs a model in production and wants to improve it without paying for a custom fine-tuning engagement. The gateway gives that company a self-serve loop. Route traffic, collect examples, fine-tune, redeploy. The improvement compounds over time. The company builds a proprietary model that is specialized to its own domain, its own users, and its own failure modes.

This is the same playbook that drove the open-source boom in the last AI cycle. Companies realized that a general model is less valuable than a model fine-tuned on their own data. The bottleneck was always the data collection. Experiential Labs removes that bottleneck by making the collection automatic. The gateway is the data pipeline in disguise.

There is a strategic angle for model providers here as well. A gateway that sits in front of multiple providers is a threat to any single provider’s lock-in. If a company can route traffic to whichever model performs best on its own evaluation set, the switching costs drop. The providers have to compete on price and quality rather than on the inertia of an existing integration. That is good for customers and uncomfortable for the incumbents.

The flywheel economics

The economic logic of the gateway is straightforward. Inference is getting cheaper every quarter. The marginal cost of a request is approaching zero. If a company can capture the signal from that request and use it to improve its model, the value of the request exceeds its cost. The gateway turns a cost center into a data asset. That is the kind of arithmetic that CFOs understand.

The open-source licensing is the other piece of the puzzle. By releasing the gateway under an open license, Experiential Labs is distributing the collection infrastructure for free. The company does not make money on the tool. It makes money on the services around it, or on the hosted version, or on the fine-tuning pipeline that sits downstream. The classic open-core model. The gateway is the funnel. The training services are the revenue.

There is a risk that the open-source release cannibalizes the commercial offering. A team with enough engineering talent could run the gateway, collect the data, and fine-tune a model without ever paying Experiential Labs a dollar. That is the tradeoff every open-source company makes. The hope is that the number of teams willing to build the full pipeline themselves is small enough that the hosted version still finds buyers.

What this means for AI builders

The deeper implication is about who owns the feedback loop. The current AI economy is built on a one-way flow. Users send requests to models, the models respond, and the interaction ends. The labs that control the frontier models get the benefit of aggregate usage patterns, but individual companies get nothing back from their own traffic. Experiential Labs is proposing a two-way flow. The traffic improves the model, and the improved model serves the traffic better. The loop is closed at the level of the individual deployment.

For builders, the practical takeaway is that the data you are already generating has value you are not capturing. Every request your application sends to a model is a labeled example of a real task. The gateway is a way to collect those examples systematically and turn them into a durable asset. The model you run today is a rental. The model you build from your own traffic is an investment.

The privacy questions are not going away. Any team that deploys this gateway needs a clear policy on what gets retained, what gets scrubbed, and what gets excluded from training entirely. The tool makes collection trivial. It does not make the governance decisions for you. The teams that treat the gateway as a fire-and-forget data collector will eventually have a compliance problem. The teams that build the governance into the pipeline from the start will have a durable advantage.

Experiential Labs has shipped a tool that makes the data flywheel accessible to any team with production traffic. The open-source release means the collection layer is now a commodity. The differentiation will come from what teams do with the data they collect, and how responsibly they handle it. The gateway is the easy part. The hard part is what comes after.