Feature Forge: turning unstructured SMB data into validated credit features
Uses an LLM offline to write deterministic feature code from long text such as shipment notes, broker emails and statements, so credit models can use those signals without calling an LLM at decision time.
Early
Written for: Small startup team
Next step
Call the head of credit risk at an inventory-finance or e-commerce revenue-based lender. Ask which unstructured fields they read by hand today and whether deterministic features built from them could pass their model validation.
Why now
SMB lenders are starting to underwrite on supply-chain data (orders, inventory, shipments), most of which is messy text. Model-risk teams accept deterministic features far more readily than live LLM calls inside a credit decision.
What became possible
LLM-BlockFE shows that an LLM can search for executable feature programs over long text offline, keeping real-time decisions on fast structured models. Cheap small models make that search affordable.
Two ways in
- Feature search workbench
Upload labelled loan outcomes and raw documents. The tool proposes, tests and versions feature code, with lift, stability and documentation for each feature.
- Customer
- Credit analytics teams at inventory-finance and revenue-based lenders
- Model
- Platform subscription plus per-model validation pack
- Supply-chain signal API
Ready-made features computed from commerce and warehouse data feeds for lenders that lack data science staff.
- Customer
- Embedded-lending programs on commerce platforms
- Model
- Per-applicant pricing
Competition
I found no commercial product that uses an LLM offline to write deterministic, versioned feature code from long-form SMB text such as shipment notes, broker emails and statements, and then tests that code against labelled loan outcomes with lift, stability and model-risk documentation. Document-AI vendors like Heron extract fields per document, and decision platforms like Taktile run LLM agents at decision time.
Full competition notes (11 competitors)
I found no commercial product that uses an LLM offline to write deterministic, versioned feature code from long-form SMB text such as shipment notes, broker emails and statements, and then tests that code against labelled loan outcomes with lift, stability and model-risk documentation. Document-AI vendors like Heron extract fields per document, and decision platforms like Taktile run LLM agents at decision time. Feature-engineering tools like FeatureByte and research like CAAFE and LLM-FE work on generic tabular data. The closest thing is an open-source Claude Agent SDK project that runs on synthetic data. So the feature search workbench aimed at credit analytics teams at inventory-finance and revenue-based lenders looks open. Its pitch would be auditable code that passes SR 11-7 style validation with no LLM in the decision path; that buyers value this is my inference. The main risk is that Heron or Taktile add the feature, and that the published method is easy to copy, so any advantage would have to come from credit-specific checks (point-in-time leakage, PSI, fair-lending proxy tests, generated validation documents) and a library of proven feature templates. The supply-chain signal API is weaker. The platforms that own fulfillment and commerce data (ShipBob with Slope, Shiprocket, Shopify) underwrite their own sellers, and Rutter and Codat already sell the raw data layer. That product looks viable only as ready-made risk features built on top of those aggregators, sold to smaller embedded-lending programs that lack data science staff.
- aleksarias/lending-agentic-underwriting (open source)
The closest match to the workbench idea. Agents research features and train challenger default-risk models built on Databricks and the Claude Agent SDK. A deterministic harness scores them on AUC/KS, lift, time and segment stability, PSI, leakage and adverse impact, and a human approves before promotion. It uses synthetic data and is a project, not a company.
- CAAFE (Context-Aware Automated Feature Engineering)
NeurIPS 2023 paper with public code. An LLM writes Python feature code from dataset descriptions, the code is executed, and features are kept only if cross-validated performance improves. It is the core technique but works on generic tabular data, not long text or credit specifically.
- LLM-FE
2025 research that adds evolutionary search to LLM-driven feature code generation for tabular data. Academic prior art, not a product.
- FeatureByte
Feature engineering platform with generative-AI feature ideation and point-in-time correct features. General-purpose ML, with no credit or unstructured-text focus found.
- Heron Data
LLM-based extraction of bank statements, applications and email submissions into structured data for SMB lenders. Raised a $16.6M Series A. It extracts per document and does not search for or validate features, but it holds the customer relationship and the document pipeline.
- Taktile
AI decision platform for lenders that launched customizable underwriting AI agents in September 2025, plus a third-party data marketplace. The LLM runs in the decision workflow rather than producing offline deterministic feature code, but it could add feature discovery.
- Slope (ShipBob financing)
Embedded lender inside ShipBob's fulfillment dashboard that underwrites on order volume, inventory turnover and shipping data. A direct example of supply-chain signals used in-house.
- Shiprocket
Indian logistics platform that uses failed deliveries and cash-on-delivery behaviour as underwriting signals for sellers.
- CapRelease
London ecommerce lender that underwrites on real-time logistics-partner, sales and payment data. Raised a combined debt and equity round in December 2025.
- Rutter
Single API for standardised commerce, accounting and billing data, used by SMB lenders such as Rapid Finance for underwriting. A data layer rather than ready-made risk features, but it sits next to the signal API idea.
- Codat
Lending data API that pulls accounting, commerce and inventory platform data (for example Brightpearl and Zoho Inventory) for SMB underwriting.
Checked 2026-10-09
Risks
- Lift over existing structured features may be small, and proving it needs historical outcome data the lender must share.
- Fair-lending review of newly generated features adds work.
- Large lenders already have data science teams.
Strong team fit; tech readiness and open space are the weak point (6).
tech 6 · pull 7 · space 6 · fit 9 · build 7
Exposed
Companies that could benefit or be threatened. News and context, not investment advice.
Momentum of its main topic
Sources
What happened next
- +7d · 16 Oct 2026 nothing yet. No launches, funding announcements, acquisitions, or coverage of Feature Forge or comparable LLM-based feature engineering products for SMB credit were found in the seven days following the flag date.
Status history
- 2026-10-09 Flagged as buildable