Ad tech
Specialists inside the RTB auction window.
The IAB OpenRTB spec allows 100 milliseconds end-to-end. Frontier API calls arrive too late to influence the bid. A domain-specialist model running on your own hardware is designed to complete brand-safety, bid-shading, and MFA classification inside the auction envelope, with no per-call API fee.
Active evidence
Current public proof is adtech-first: Assay-Adtech v1 is the active released benchmark surface.
Buyer constraint
The decision must clear before the bid expires.
Claude, ChatGPT, and Gemini are useful for analyst workflows, but a remote frontier call misses the auction path and adds a variable per-decision bill.
A small specialist runs next to the bidder, scores only the bounded auction task, and promotes only after held-out RTB scenarios clear.
100ms auction budget
Specialist on controlled hardware completes inside the envelope.
A frontier API is still useful as a benchmark, not as the bid-path actuator.
Evidence posture
Assay-Adtech v1
Public benchmark evidence is adtech-first, with released Assay-Adtech v1 artefacts covering brand safety, MFA detection, bid shading, and RTB-style evaluation.
Published on Agentsia Labs
Smith note
Smith is watching the evidence posture: adtech has the active public proof today. Candidate domains stay labelled as future work until their own rubrics, datasets, and results are ready.
Deployment topology
T1 is the default pattern for ad tech.
- T1
Co-located training and inference
The same customer-controlled hardware can train the specialist and serve the bounded workflow.
Recommended for this vertical
- T2
Trained centrally, deployed to the edge
The signed artefact is trained centrally, then shipped to vehicle, mobile, embedded, or edge hardware.
- T3
Trained centrally, deployed to a commercial runtime
Designed for customer-hardware training with a signed specialist served through a managed inference vendor when the customer does not want to run its own inference layer.
Why a specialist model
Three reasons frontier models do not fit.
- Frontier API latency exceeds the auction window
- The IAB OpenRTB spec allows 100ms end-to-end. A frontier API round-trip often adds hundreds of milliseconds before a response arrives. A small specialist model on your own hardware fits the auction envelope.
- Per-decision cost is prohibitive at auction scale
- Running hundreds of millions of daily bid requests through a frontier API at commercial rates costs more than the CPM on most inventory. A specialist on owned hardware converts a variable per-token cost into controlled infrastructure cost with no per-call API fee.
- Bid logic must not leave your network
- Your blocklist weights, brand-safety thresholds, and shading coefficients are core IP. Sending bid-request payloads to a third-party API exposes that logic and creates a data-egress record that complicates buyer and publisher contracts.
Use cases
Concrete workflows, not a category claim.
Each use case below maps to a real workflow a design-partner team would bring to Modelsmith. The specialist model can be trained on your data, evaluated against your rubric, and promoted through your governance path where that operating loop is enabled.
Brand-safety classification
Train a specialist to score inventory against your IAB-based blocklist using page context, URL, and creative-adjacency signals. Human reviewers set the thresholds. The model enforces them at scale without per-call cost.
Evidence scenarios
Bid shading
Fine-tune a specialist on your historical win and loss data to predict the optimal clearing price below the first-price ceiling. The model adapts to shifting auction dynamics rather than reading from a static lookup table.
Pre-bid MFA filtering
Classify made-for-advertising inventory before the bid is placed. Train on your own MFA-labelled dataset so the model reflects your definition of quality, not a third-party taxonomy imposed at the exchange.
On-device gaming SDK
A sub-4B specialist embedded in a mobile gaming SDK for contextual ad placement. No network call required. Classification runs on the device with no user data leaving the handset.
Where it does not fit
A specialist is the wrong answer unless the workflow is bounded.
The strongest buyers know what they are trying to control: latency, data movement, auditability, model size, or edge deployment. If the work is broad, casual, or unconstrained, a frontier-lab model is usually the simpler answer.
A creative generator for ad copy or campaign names.
A DSP, SSP, ad server, or attribution platform replacement.
A broad marketing copilot where latency and bid-path control do not matter.
Get started
Apply for the adtech design-partner cohort.
Apply with a real adtech workflow in mind. We will scope the Synthetic POC together, run a complete specialisation cycle on synthetic domain scenarios, and hand you a validated model with a full evidence bundle before any licence commitment.