Core technology & R&D
Algorithms, then models, then agents — layer on layer
We have built a complete body of algorithms and technology in large models, efficient model production, AI anti-fraud and computer vision, with sustained R&D in industry-model distillation, standardized modeling, product design, intelligent marketing, digital operations and agents. It all comes together in the SAM Data Analysis & Scorecard Modeling Platform, serving banking, FMCG and marketing, pharmaceuticals and other industries.
The most error-prone parts of modeling, rewritten as engineering constraints
SAM earns trust through four engineering designs, not through model good behavior.
Every statistic is computed by code in a signed image
WOE, IV, KS, VIF, PSI and other statistics come from deterministic code; the language model only explains and proposes candidates, and any figure it “works out” is never accepted as evidence.
- Same data, same parameters, same result — every time
- Every number traces back to a code version and data snapshot
Change one thing, re-run only what it touches
The fourteen modeling steps form a directed acyclic graph: change any parameter, snapshot or binning and every downstream step — with the approvals it held — is voided. Only affected steps re-run; earlier runs are preserved in full.
- The blast radius of a change is computed automatically
- Voided approvals must be signed again
Validation is separate from development, and release has one gate
SAM-05 validates in a separate workspace, with a separate sample and signing key, and development can never rewrite its conclusions; the Release Controller is the only way any artifact reaches production.
- A named signatory at each of fourteen gates
- Timeouts suspend — nothing passes by default
Runs in your own environment, on modest compute
Inside the controlled modeling sandbox, work gets done, data can’t leave and records can’t be altered; every release produces an immutable run manifest that can replay the whole modeling run.
- Data and artifacts never leave the sandbox
- The manifest binds data, code, parameters, sign-offs, model and monitoring plan
Years of accumulated technology and data
Samples, variables, tools, products — a closed loop
Our consistent results come from four capabilities built up across the entire modeling chain.
Modeling sample handling
Interest-only, amortizing and irregular loans are handled separately; full, bad, good and indeterminate samples are strictly defined for each definition; and no risk margin is allowed to creep into modeling samples.
Derived-variable design
Multi-dimensional “time + product + status + value + statistic” design across personal credit reports, tax, invoices, bank statements, scenario and third-party financial data has produced 20,000+ interpretable variables with proven KS — now preloaded in the VDC platform. The same five-dimension method applies to detailed records in FMCG, pharmaceuticals and other industries.
The variables quantify both ability and willingness to repay, and trace fine-grained changes in borrower behavior before, during and after the loan, and in marketing response, across fifteen dimensions.
Modeling tools
Our modeling tools descend from those used for 30 years at JPMorgan’s headquarters, refined over three years in China by seven PhDs and more than ten master’s researchers: practiced and standardized on 10 million+ samples, covering 3,000+ modeling details, extended with 100+ proprietary statistical and machine-learning algorithms, and validated across 100+ projects.
With this toolkit we have competed head-to-head with 200 teams from 60+ vendors and banks, consistently leading on discrimination, model and strategy stability, rank ordering and performance.
Product design
We have designed 30+ scenario products in China and built a complete methodology for product elements, business processes and risk frameworks, resolving hundreds of real-world problems through iteration.
Coverage includes marketing and risk whitelists, anti-fraud, admission rules, master scorecards, limits and pricing, owner and general due diligence, approval, compliance, drawdown models and strategies, post-loan early warning, drawdown-rate, renewal and revolving-credit models, collection strategies, asset preservation, ABS strategies, customer tiering, personalized repayment, primary and secondary guarantees, revenue verification and risk-limit trade-offs.
Four areas, fifteen capabilities
Computer vision is a core part of our technology, spanning biometrics, vehicle intelligence, document intelligence and image processing.
Biometrics4
- Face recognition
- Liveness detection
- Cattle face recognition
- Body length & weight estimation
Vehicle intelligence4
- License-plate recognition
- Vehicle model recognition
- Part segmentation
- Damage detection
Document intelligence4
- Bill extraction
- Text detection & recognition
- Table-line recognition
- ID & license OCR
Image processing3
- Distortion correction
- Foreground segmentation
- Image quality assessment
Recognition accuracy
We use the deep-learning RCNN algorithm for facial feature extraction and matching, and biosignal detection (PPG analysis) for liveness — effective against photo, video and 3D-mask attacks.
DFA, our in-house distributed ML platform
Built on Hadoop, Yarn and Docker, DFA spans four layers from the bottom up, providing distributed storage, permission management and distributed training.
- InteractionWeb Portal
- ApplicationDFLabeling · DFLearning · DF-SDK
- PlatformYarn · HDFS + HUE
- HardwareGPU compute servers
Legal and technical assurance
In computer vision alone we hold 31 software copyrights and 12 patents. Together with our proprietary algorithms and modeling tools, they form a complete IP portfolio that protects continuous product iteration and on-premises delivery.
Every SAM feature, online, for up to 10,000 samples
Upload up to 10,000 records and run the whole fourteen-step pipeline — data contract to packaged scorecard, human gates included — to see for yourself whether SAM fits your business.
Start with one business question. Let SAM deliver a model you can sign off.
Whether you are in banking, FMCG and marketing, pharmaceuticals or another industry, try SAM online first — or talk to our modeling experts about your use case.