⊟
CRM field normalizer
RevOpsCleans and standardizes messy Salesforce account/contact rows into a deterministic schema — canonical company name, normalized industry, employee band, compliance-framework tags, a dedupe key, and a data-quality flag — so the whole agent fleet reads from one clean hub.
The routing recipe
Job typeStructured extractionPolicyCheapest that passes
routes to
DiffusionGemma 26B-A4B nvidia
nvidia:google/diffusiongemma-26b-a4b-it
92.5
match score
78.5
task fit
100
cost
97.7
speed
Why this model from benchmarks & capabilities
- "Structured extraction" leans hardest on instruction; DiffusionGemma 26B-A4B scores 76/100 there — #40 of 54.
- Strength on this class of work shows up on IFEval, IFBench (instruction following), where Google models are competitive.
- Cost: runs on NVIDIA free credits / self-hostable NIM — effectively $0 marginal. Speed: ~1100 tok/s (estimated).
- Chosen over Nemotron 3.5 Lightning 30B-A3B for faster.
| Capability the job needs | Weight | DiffusionGemma 26B-A4B | Catalog rank |
|---|---|---|---|
| instruction | 1.00 | #40 of 54 | |
| tool_use | 0.60 | #47 of 54 | |
| speed | 0.60 | #1 of 54 |
Relevant benchmarksIFEval IFBench (instruction following)
EconomicsCost: runs on NVIDIA free credits / self-hostable NIM — effectively $0 marginal.
Speed: ~1100 tok/s (estimated). vs runner-upChosen over Nemotron 3.5 Lightning 30B-A3B for faster. (Nemotron 3.5 Lightning 30B-A3B, score 92.4)
Speed: ~1100 tok/s (estimated). vs runner-upChosen over Nemotron 3.5 Lightning 30B-A3B for faster. (Nemotron 3.5 Lightning 30B-A3B, score 92.4)
Example result
Sample input ▾
Raw SFDC row → Account Name: "acme payments inc." | Website: "https://www.AcmePay.io/pricing" | Industry: "financial / payments saas" | Employees: "~85" | Country: "USA" | Notes: "pursuing SOC2 type 2 + PCI for enterprise deal, asked about ISO too" | Billing State: "CA"
{
"canonical_name": "Acme Payments",
"domain": "acmepay.io",
"industry": "FinTech",
"employee_band": "51-200",
"country": "US",
"compliance_frameworks": ["SOC2", "PCI"],
"data_quality": "needs_review",
"issues": [
"Name: stripped 'inc.' and title-cased 'acme payments inc.' -> 'Acme Payments'",
"Industry: 'financial / payments saas' resolved to FinTech (payments signal outweighs generic 'saas')",
"Headcount: '~85' is approximate -> bucketed 51-200; raw value preserved in audit field",
"Domain: stripped scheme + www + '/pricing' path -> root 'acmepay.io'",
"ISO27001: only 'asked about ISO too' — a soft inquiry, NOT a commitment; excluded from compliance_frameworks and flagged for SDR follow-up"
],
"dedupe_key": "acmepay.io"
}
Test it on your own data