Fraud detection & prevention
Risk signals, provider and transaction classification, anomaly detection and evidence-led investigations.
AI RESEARCH / FINANCE / INSURANCE
EST. 2022AI research in finance.
State-of-the-art AI, applied to your organisation — from peer-reviewed methods to auditable systems that work inside real financial workflows.
Applied R&D · London, United Kingdom
ARIFIN / LAB NOTE 014MODEL → SYSTEM
01Research a measurable financial decision
02Build with provenance, controls and evidence
03Deploy into a human-owned workflow
ACCOUNTABLE BY DESIGNHUMAN / LOOP
01 / SERVICES
Applied, not ornamental.
We translate current AI research into systems your organisation can operate, inspect and improve — without losing the thread between a model result and the decision it shapes.
Predictive modelling for finance
Decision models for underwriting, forecasting, pricing and portfolio workflows.
Intelligent document processing
Turn policy, claim and financial documents into structured, traceable data.
AI agents & workflow automation
Goal-directed agents with bounded tools, approval gates and complete operational records.
Model audit, explainability & governance
Make model behaviour legible to operators, reviewers, regulators and affected people.
Applied R&D partnerships
Joint research, prototypes and field evaluation for difficult financial problems.
02 / PAPER LIBRARY
Research worth implementing.
A focused reading shelf across fraud detection, asset pricing, quantitative trading and financial language models. Every third-party paper in the library is paired with its official implementation; our own featured research has a live implementation on this page.
Improvement of decision tree classifier accuracy for healthcare insurance fraud prediction by using Extreme Gradient Boosting algorithm
ICIMCIS 2020. Healthcare insurance fraud classification with an accuracy lift from Random Forest to XGBoost.
Hire us to implement itXGBoost: A Scalable Tree Boosting System
KDD 2016. The system paper behind a widely used gradient-boosted tree library.
Hire us to implement itDeep Learning in Asset Pricing
Management Science. A deep-learning approach to asset-pricing models and economic structure.
Hire us to implement itFinRL: A Deep Reinforcement Learning Library for Automated Stock Trading in Quantitative Finance
NeurIPS 2020 Deep RL Workshop. An open-source framework for reproducible financial reinforcement learning.
Hire us to implement itFinGPT: Open-Source Financial Large Language Models
Open financial language models, data pipelines and evaluation resources for research and applications.
Hire us to implement itQlib: An AI-oriented Quantitative Investment Platform
A Microsoft open-source platform for quantitative investment research, model training and strategy workflows.
Hire us to implement itIMPLEMENTATION FLOW
From paper to production.
After you contact us, we turn a relevant research result into a measured system built around your data, constraints and operating environment.
Contact
You tell us the problem via the contact section below.
Discovery
We map your data, constraints and success metrics.
Research selection
We pick the right paper and method for your case.
Pilot implementation
A working prototype on your own data.
Evaluation
Measured against your baseline, reported honestly.
Production
Hardened deployment, monitoring and handover with docs.
03 / FEATURED RESEARCH
Five points of accuracy, earned.
The study used Kaggle Medicare provider-fraud data covering medical beneficiaries and claims. Random subsampling with Extreme Gradient Boosting improved overall fraud-provider classification accuracy from an 81% Random Forest baseline to 86%.
N. A. Akbar, A. Sunyoto, M. Rudyanto Arief & W. Caesarendra. ICIMCIS, Jakarta, November 2020, pp. 110–114. DOI / IEEE Xplore ↗
Accuracy comparison
OVERALL ACCURACYAbsolute accuracy lift+5 pts
01 / DATAMedicare beneficiaries + claims
02 / BASELINEDecision-tree ensemble / Random Forest
03 / METHODXGBoost with random subsampling
04 / OUTPUTFraud-provider classification
Reflex — autonomous deal-watcher agent
An entry for the NVIDIA Claw Agent Challenge (Spain edition): an ID→EU arbitrage deal-watcher claw agent with a personal System One decision model and a working web demo.
04 / FUTURE SHOWCASE
Provable claims — like git, for the public.
Most people lose the path through a claim not because their case is wrong, but because the rules, evidence requirements and appeal rights are hard to see. Smart Claims turns each step into a readable, verifiable record, while people keep authority over every external action.
THE RULES, FINALLY VISIBLESmart Claims workspace
LOCAL SIMULATION · NOT SUBMITTEDFile a mock claim
Use fictional details only. Nothing is saved or sent from this demo.
AI intake agent
The agent checks the packet, maps the insurer’s reason to plain-language requirements and exposes the next available action.
Claim form
Required identity and treatment fields
FOUNDRequired identity and treatment fields
Clinical note
Evidence referenced by denial reason
MATCHEDEvidence referenced by denial reason
Right to appeal
Internal appeal is available before escalation
VISIBLEInternal appeal is available before escalation
Plain-English rule match
“Provide the clinical note and ask for an internal review of the denial.”
“Provide the clinical note and ask for an internal review of the denial.”
Hash-chained public ledger
Each event includes the previous event’s hash. Change one record and the chain no longer verifies.
READY TO VERIFY
Email follow-up agent
The agent can prepare the case, but a person must approve before the simulated send.
TO: Claims review teamRequest for internal review — mock claim
I’m requesting an internal review of this claim denial. The clinical note referenced by the denial reason is attached to the claim record. Please confirm the applicable policy clause, the review timetable, and any remaining evidence required.
HUMAN APPROVAL REQUIRED
Simulation only — no email leaves this page.
Voice agent, human-gated
Approve a mock call and watch the agent follow the same evidence record without inventing new facts.
AGENT I’m calling about a mock claim currently under internal review.
REVIEWER What evidence has been added?
AGENT The clinical note referenced by the denial reason is recorded in the evidence ledger.
AGENT Please confirm the governing policy clause and decision timetable.
HUMAN APPROVAL REQUIRED
Simulation only — no call is placed.
Decision with a complete audit trail
The system does not promise an outcome. It makes the case, evidence and next human decision inspectable.
Appeal packet ready for human review
- Denial reason mapped to required evidence
- Clinical note recorded in the ledger
- Email and voice follow-up approvals recorded
- Hash chain available for verification
NOT YET VERIFIED
05 / WHY THIS MATTERS
Denial is often the start of a process. People are not told that.
Claim friction harms individuals and providers. These figures are dated, sourced snapshots — not claims about every market or every insurer.
19%of in-network health-insurance claims were deniedUS ACA MARKETPLACE · 2024 · KFF ↗
<1% / 1 in 3Fewer than 1% of denied claims are appealed — yet about 1 in 3 internal appeals gets overturnedKFF · 2024 ↗
40%of consumers knew they had a legal right to appeal a denialKFF ↗
Rp19tnBPJS Kesehatan unpaid claim obligations to Indonesian hospitalsOCT 2019 · BPJS DIRECTOR STATEMENT ↗
439 hospitalsin East Java lost a combined Rp500bn to 12,000 pending or disputed BPJS claims2024 · PERSI EAST JAVA ↗
Rp78–100bnin unpaid BPJS claims at RSUD Ulin Banjarmasin, with roughly one month of cash left — “terancam bangkrut”2019 ↗
1,322insurance complaints to Indonesia’s OJK, driven by slow claim processesJAN–NOV 2024 ↗
Hospitals pushed to the brink — not a claim of closures. The cited evidence documents severe unpaid and disputed claims, losses and liquidity pressure. We found no verified case establishing that a hospital closed or was sold because of unpaid BPJS claims.
06 / CONTACT
London W12 7TA
United Kingdom [email protected]
Make the next decision legible.
80 Wood LnLondon W12 7TA
United Kingdom [email protected]
80 Wood Ln, London W12 7TA · Map data legal notices available in the map control.