The IRIS Sales Model is IRISeller's proprietary Sales Language Model (SLM) — a compact, domain-specific AI system trained from the ground up exclusively on B2B sales data. Unlike general-purpose large language models that attempt to understand everything, IRIS is architected for precision: understanding, classifying, and acting on the language of sales with a depth no off-the-shelf model can replicate.
It analyses natural language inputs from sales conversations, emails, CRM notes, and deal records — returning structured, machine-readable intelligence across six critical sales dimensions in under 40 milliseconds. The model has been trained to recognise the subtle signals embedded in everyday sales language: the difference between a prospect who is stalling and one who is genuinely re-evaluating, between a compliance objection and an integration concern, between a reschedule and a rejection.
The result is a classification engine that behaves like an experienced sales professional reading between the lines — at API speed, at scale, without human effort.
Why a purpose-built SLM? General-purpose large language models — despite their scale — are not well-suited for high-throughput classification in production. They carry billions of parameters of irrelevant general knowledge, require expensive inference infrastructure, and deliver inconsistent probabilistic outputs that are difficult to trust in business-critical pipelines.
The IRIS SLM is architected differently. Its entire parameter budget is dedicated to one domain. Every layer of the model — from tokenisation to attention heads to the classification head — has been calibrated on B2B sales language. The model is compact enough to serve thousands of requests per minute on a single GPU, yet deep enough to handle the full complexity of real-world sales communication: informal phrasing, missing context, domain jargon, and mixed-register writing.
This is not a fine-tuned chatbot. IRIS is a discriminative classification model — it does not generate, it decides. And it decides with the confidence of a system that has seen over 120,000 labelled examples of exactly the language it will encounter in production.
How it works: Submit any freeform sales text — an email body, a call transcript excerpt, a CRM note, or a deal context paragraph — to any of the six classification endpoints. The model tokenises the input, passes it through all transformer layers, and produces a softmax probability distribution over every possible class. The predicted label is the highest-probability class; the confidence score reflects how decisively the model selected it. The full probability distribution is returned so downstream systems can apply their own thresholds or build ensemble logic.
The foundation of the IRIS Sales Model is a purpose-built, proprietary dataset constructed entirely from the ground up. IRISeller's research team developed a comprehensive ontology of B2B sales language — mapping the vocabulary, speech patterns, and contextual signals that define each stage of the modern sales cycle.
Over 120,000 labelled examples were generated across six classification tasks, spanning diverse industries, company sizes, seniority levels, and deal contexts. Each example was engineered to reflect the real-world variability of sales communication — varying in formality, length, terminology, and intent signal strength.
To build a model that generalises beyond clean, well-formed text, every dataset was subjected to IRISeller's proprietary augmentation pipeline. This process deliberately introduced the linguistic imperfections and stylistic variations that characterise real-world sales communication — typos, abbreviations, informal phrasing, punctuation variations, and mixed-register writing styles.
40% of all training examples were augmented at varying intensity levels — from subtle character-level perturbations to significant structural rewrites — ensuring the model is robust to the full spectrum of real-world input quality.
The IRIS Sales Model is architected as a Sales Language Model (SLM) — a compact, bidirectional transformer trained specifically for the B2B sales domain rather than general language understanding. The model's architecture is optimised for classification tasks: it builds deep contextual representations of sales text without the generative overhead of large-scale foundation models.
In its domain adaptation stage, the model's internal representations are fully recalibrated to the vocabulary, intent signals, and semantic patterns unique to sales. General-purpose models treat phrases like "the deal is getting cold", "we need to loop in procurement", or "our champion just left the company" as ordinary language. The IRIS SLM has been trained to recognise these as high-signal sales indicators — with precision and confidence no off-the-shelf model can match.
Six dedicated classification heads were trained independently on their respective datasets — each optimised with task-specific hyperparameters, learning schedules, and sequence length configurations suited to the nature of its inputs. Lead scoring handles long-form prospect profiles; intent detection operates on short queries; deal risk assessment processes multi-sentence deal notes.
Each classifier was trained with cosine learning rate scheduling, gradient accumulation, and dynamic padding to maximise sample efficiency and convergence stability. Early stopping with a patience threshold ensured each model was preserved at its point of peak generalisation — not over-fitted to training data.
Every classifier was evaluated on a held-out test set that was completely isolated from training — never seen during model development. Performance was measured not just by accuracy but by macro F1 score, Matthews Correlation Coefficient (MCC), per-class precision and recall, and confusion matrix analysis — ensuring no individual label category was underserved.
Results were benchmarked against majority-class and stratified random baselines to quantify genuine model intelligence above statistical shortcuts. The IRIS Sales Model outperforms all baselines by a substantial margin across every task, confirming that it has learned true language understanding — not surface-level pattern matching.
Following validation, each IRIS SLM classifier underwent weight consolidation — all domain adaptations and task-specific learning merged into a single, unified set of parameters. This produces a deployment-optimised model that requires no additional runtime dependencies, loads directly into GPU memory, and serves predictions with sub-40ms latency.
The serving layer is designed for reliability and consistency: models are loaded once at startup, pinned to GPU memory, and served through an async inference engine that handles concurrent requests without queuing delays. The result is a production system capable of classifying thousands of sales signals per minute with stable, predictable performance.
Evaluates a prospect's buying signals, engagement level, and profile to classify leads into actionable tiers, helping sales teams prioritise outreach and resource allocation.
Identifies the underlying intent behind sales representative queries and requests, enabling intelligent routing, automation triggers, and CRM workflow activation.
Measures the emotional tone of prospect communications across five granular sentiment levels, enabling sales teams to tailor responses and identify at-risk relationships before they deteriorate.
Identifies the root category of a prospect's objection, enabling sales representatives to access the right playbook, talking points, and case studies to address the specific barrier effectively.
Assesses the health of an active deal by analysing contextual signals — stakeholder changes, timeline slippage, competitive pressure, and engagement patterns — to surface deals requiring immediate attention.
Automatically categorises inbound sales emails to power intelligent routing, automated responses, and CRM record enrichment — eliminating manual triage and ensuring high-priority leads receive immediate attention.
Automatically enrich CRM records with lead scores, email categories, and sentiment scores as new data arrives — without any manual input from reps.
Trigger real-time Slack or email alerts when the Deal Risk model detects critical or high-risk signals in deal notes, ensuring managers can intervene early.
Route hot leads instantly to senior reps and warm leads into nurture sequences — powered by real-time Lead Scoring classification on every new inbound.
Analyse call transcripts and chat logs in real time to surface objections, detect sentiment shifts, and recommend the right response playbook mid-conversation.
Automatically categorise every inbound sales email and trigger the right workflow — immediate callback for meeting requests, nurture for info requests, discard for spam.
Aggregate classification outputs across your entire pipeline to generate intent heatmaps, objection frequency reports, and sentiment trend dashboards.
All API requests must include an X-API-Key header. The base URL for all endpoints is https://model.iriseller.com. Responses are returned as JSON.
Public health check. Returns model status and number of classifiers loaded. No authentication required.
Classify a single text input. Replace {task} with one of: lead_scoring, intent_detection, sentiment_analysis, objection_classification, deal_risk_assessment, email_categorization.
Classify up to 100 texts in a single request. More efficient than individual calls for bulk processing.