Reperks
AI agent Perks eliminates 98% of landlord billing work
Industry
Real Estate / Property TechnologyProject type
AI Agent Development, Document Processing AutomationProject duration
6 weeks MVP + Ongoing DevelopmentTeam size
5 specialists
At a glance
What Perks delivers for landlords:
Time savings
Complete annual billing in 15 minutes instead of 20+ hours (98% reduction)Legal protection
Automatic compliance with German rental law eliminates costly disputesZero manual entry
Upload documents in any format. Perks extracts and validates everythingStress-free accuracy
Triple-validated calculations guarantee mathematical precisionRapid deployment
Production-ready system delivered in 6-week MVP sprint
What the client says
"The team has been highly engaged, proactive, responsive, and available to assist, and has participated actively in reviews and retrospectives. The team's collaborative and solution-oriented approach stands out."
Technical Founders
Reperks
About the client
Reperks is a German PropTech startup targeting the EUR 400+ billion German rental market. Their focus: digitizing the manual processes that cost landlords hundreds of hours every year.
HyperSense first built Reperks a classical MVP for utility billing. Solid engineering, clean product. The market didn't bite. So we pivoted and rebuilt the core as an AI agent. That pivot changed everything. The product is now live at agent.reperks.de and growing rapidly.
Reperks' vision extends beyond billing to a full suite of AI-powered tools for rental property management.
Founded
Headquarters
Market
Problem space
The challenge
Market context
German landlords must create legally compliant utility cost billing statements ("Nebenkostenabrechnung") annually for each tenant. This process is notoriously complex. It involves multiple legal frameworks (BGB §556-560, HeizKV, CO2KostAufG), precise mathematical allocations based on different distribution keys (area, consumption, persons), and strict formatting requirements. Errors can result in legal disputes, lost revenue, or expensive property manager fees (EUR 50-150 per unit annually).
Current tools force landlords into rigid, form-based interfaces that don't match their actual workflows, require extensive manual data entry, and fail to provide intelligent guidance through the legal complexities.
The technical and business problems
AI reliability gap
Initial prototype using Gemini 2.5 Pro achieved only 75% accuracy with a 95-page monolithic prompt that was unmaintainable and unpredictableDocument processing complexity
Landlords provide wildly inconsistent documents (handwritten notes, poor-quality scans, property manager summaries, individual invoices) requiring intelligent OCR with confidence scoringLegal compliance risk
Germany's rental law is complex and unforgiving. Even small calculation errors or missing legal disclosures can invalidate entire billing statementsDual workflow challenge
System must support both landlords with property management (summary documents) and self-managing landlords (20+ individual invoices)Human-in-the-loop validation
Can't fully automate due to document quality issues. Must integrate user confirmation without breaking conversational flow
Additional constraints
Tight timeline
Q1 2026 launch target due to annual billing cycle (June-December prime season)Cost control
AI token costs can spiral fast. Needed efficient orchestration and caching strategiesRegulatory precision
Zero tolerance for mathematical errors. Must implement deterministic calculation validation alongside LLM processingMarket education
Target users (50+ year-old landlords) unfamiliar with conversational AI interfacesScalability requirements
Architecture must support 10,000+ concurrent users during peak season without degraded performance
The strategic goal
Give landlords back their time by transforming 20+ hours of annual billing drudgery into a 15-minute conversation with Perks, so they can focus on growing their portfolio while Perks handles legal compliance and mathematical precision automatically.
Our approach
Why HyperSense
Reperks chose HyperSense because production AI agents require architectural expertise beyond prompt engineering. We brought:
We brought:
- AI engineering depth: Agentic workflows, multi-model orchestration on AWS Bedrock
- Regulated industry experience: Compliance-focused systems where mathematical precision is non-negotiable
- Rapid MVP delivery: Production-grade systems in 6-week sprints
- AWS funding: As an AWS Partner, we brought co-funding to the project, reducing Reperks' investment while accessing enterprise-grade infrastructure
How we built it
Discovery and architecture design (Week 1)
We began with intensive analysis of the client's 95-page prompt, existing user workflows, and German legal requirements. Rather than simply refactoring the prompt, we designed a proper agentic architecture separating concerns: document processing, conversational orchestration, calculation engines, and legal validation. This allowed us to optimize each component independently and implement proper testing.
Iterative development with client validation (Weeks 2-5)
Using AWS Bedrock's flexible model routing, we're building Perks as a Claude 4.5 Sonnet-based agent with Extended Thinking capability, allowing it to reason through complex legal edge cases. We developed a dual-repository structure (frontend/backend) with Infrastructure as Code (Terraform) for rapid iteration. Reperks participates in weekly demos, providing real-world documents and landlord feedback that shapes the human-in-the-loop confirmation flows.
Legal and mathematical hardening (Week 6)
LLMs can fail at complex arithmetic. We moved all calculation logic into deterministic code with triple-validation (three independent calculation methods cross-checked). We're integrating German legal knowledge bases (BGB, HeizKV, CO2KostAufG) as structured context rather than raw prompt text, enabling Perks to provide precise legal citation and compliance checking.
What we're delivering
Perks: The AI agent that landlords actually enjoy using
Conversational billing interface
Conversational interface that guides landlords through billing in natural language. No forms, no frustrationIntelligent document processor
Intelligent document processor that reads any format (handwritten notes, scans, PDFs) and extracts data automaticallySmart workflow detection
Smart workflow detection that adapts to each landlord's situation (property manager or self-managed)Built-in legal expert
Built-in legal expert that validates compliance with German rental law at every stepError-proof calculation engine
Error-proof calculation engine that triple-checks every number so landlords never worry about mistakesProfessional PDF generation
Professional PDF generator that produces legally compliant statements tenants can't dispute
Supporting infrastructure:
AWS Bedrock integration
AWS Bedrock integration with Claude 4.5 Sonnet (Extended Thinking)Scalable serverless architecture
Scalable serverless architecture (Lambda, API Gateway, S3)PostgreSQL database
PostgreSQL database with property/tenant context persistenceReact conversational UI
React-based conversational UI with real-time document uploadAWS Cognito authentication
AWS Cognito authentication with landlord/property manager segmentationCloudWatch monitoring
CloudWatch monitoring with token usage tracking and cost alertingInfrastructure as Code
Infrastructure as Code (Terraform) for reproducible deployments
Architecture and technology
Our architecture prioritizes reliability, cost-efficiency, and maintainability over premature optimization. We designed for the MVP constraint while enabling evolution toward a multi-agent system as the product matures.
AI/ML
Technologies
AWS Bedrock, Claude 4.5 Sonnet (Extended Thinking)
Purpose
Conversational orchestration, document understanding, legal reasoning. Extended Thinking enables complex multi-step legal analysis.
Document Processing
Technologies
AWS Textract, Bedrock Data Automation
Purpose
OCR with confidence scoring for handwritten and low-quality scans. Textract for structured extraction, Bedrock for intelligent correction.
Backend
Technologies
Node.js, Express.js, PostgreSQL
Purpose
RESTful API with conversation state management, calculation engines, and property context persistence. PostgreSQL for ACID-compliant audit trails.
Frontend
Technologies
React, Axios
Purpose
Conversational UI with document upload, real-time feedback, and data confirmation flows. Mobile-responsive for on-the-go landlord access.
Infrastructure
Technologies
AWS Lambda, S3, CloudWatch, Terraform
Purpose
Serverless for cost-efficiency and auto-scaling during peak season. Terraform for reproducible infrastructure deployments.
Authentication
Technologies
AWS Cognito
Purpose
Multi-tenant authentication with landlord/property manager segmentation for future pricing tiers.
Monitoring
Technologies
CloudWatch, Mattermost
Purpose
Real-time alerting for critical errors, token usage spike detection, and system health monitoring.
Technical highlights
Agentic orchestration with cost controls
Rather than relying on heavyweight frameworks (LangChain), we built a lightweight orchestration layer that routes conversations through specialized processing nodes based on intent detection. This gives us fine-grained control over token usage and enables request-level caching for repeated legal queries.
Deterministic math + LLM reasoning
All financial calculations run through three independent validation methods (percentage-based, unit-based, reverse-calculation) in hardened code, while the LLM handles conversational guidance and legal explanation. This hybrid approach eliminates mathematical hallucinations while preserving the natural language interface users need.
Document classification and templated extraction
We implemented a two-phase document processing pipeline. First, Claude classifies the document type (invoice, property manager summary, rental contract, meter reading). Second, type-specific extraction templates optimize OCR accuracy and validation rules. This dramatically improved extraction reliability for non-standard documents.
Key features and capabilities
Upload any document, Perks handles the rest
What landlords get
Simply take photos of invoices, contracts, or statements with your phone. Perks reads everything automatically, even handwritten notes or faded faxes. You review and confirm the extracted data in seconds, then move on.Why this matters
You'll never spend hours retyping invoice data again. Perks eliminates 100% of manual data entry while keeping you in control. You see exactly what was extracted and confirm it's correct before proceeding. You save time without sacrificing accuracy or peace of mind.Technical approach
AWS Bedrock Data Automation for initial OCR with confidence scoring, followed by Claude-powered extraction against document-type-specific schemas. Extractions below 98% confidence trigger user confirmation flows with highlighted fields. Confirmed data is tokenized and standardized before database insertion to enable year-over-year reuse.
Perks adapts to your workflow, not the other way around
What landlords get
Whether you have a property manager or handle everything yourself, Perks automatically detects your situation from the first document you upload and guides you through the exact steps you need. No irrelevant questions, no wasted time.Why this matters
You're not forced into a one-size-fits-all process that doesn't match your reality. Property manager clients get a 5-minute experience verifying summary documents. Self-managing landlords get intelligent guidance organizing 20+ invoices. Either way, you complete billing faster because Perks understands your unique workflow.Technical approach
Intent classification on initial document upload triggers pathway selection. For property manager pathway, agent validates summary document against legal requirements and prompts for missing government invoices. For self-managed pathway, agent orchestrates multi-document collection, validates completeness across required categories, and guides allocation calculations. Both converge to the same final billing generation endpoint.
Sleep better with built-in legal protection
What landlords get
Perks acts as your personal legal expert, continuously checking every piece of data against German rental law (BGB, HeizKV, CO2KostAufG). If something's wrong, you'll know immediately. Before you send anything to tenants.Why this matters
You avoid costly legal disputes and tenant payment withholding. One missing disclosure or incorrect heating cost split can invalidate your entire billing and tie you up in months of legal headaches. Perks catches these errors automatically, so you can send billing statements with complete confidence that tenants can't challenge them.Technical approach
Legal knowledge bases (structured as JSON schemas with rule definitions) are loaded into Claude's context using RAG. For each data point (e.g., cost position, allocation key, rental contract clause), the agent cross-references legal requirements and flags violations. Final validation runs a comprehensive checklist before PDF generation, blocking output until all requirements are satisfied.
Never worry about math mistakes again
What landlords get
Every euro, every allocation, every percentage. Calculated three different ways automatically and cross-checked for perfect accuracy. If the numbers don't match down to the cent, Perks won't let you proceed.Why this matters
You'll never lose sleep wondering if you made a calculation error that could cost thousands in disputes. Mathematical mistakes destroy tenant trust and create legal liability. Perks' triple-validation system guarantees perfect accuracy every time, so you can focus on landlord-tenant relationships instead of spreadsheet formulas.Technical approach
All calculations are moved outside the LLM into hardened TypeScript/Node.js functions. Each allocation runs through: (1) percentage-based calculation, (2) unit-based calculation, (3) reverse-validation (sum of all tenants = total cost). Results are compared; mismatches trigger error logs and block progression. The LLM receives calculation results as structured data and focuses solely on explaining them in natural language.
Professional billing statements your tenants can't dispute
What landlords get
Two versions of every billing automatically: a quick-scan overview so you can review in 2 minutes, plus a formal, print-ready PDF that meets every German legal requirement. Perfectly formatted, completely professional.Why this matters
You save hours on formatting and legal compliance research. The professional PDF includes everything German law requires: objection deadlines, tax-deductible classifications, itemized breakdowns. Your tenants receive bulletproof documentation they can trust, which means fewer disputes and faster payment collection.Technical approach
LaTeX-based PDF generation engine with template injection. Quick overview renders as Markdown tables in-app. Full legal PDF includes property/tenant headers, structured cost tables with allocation key explanations, heating cost distributions (HeizKV compliant), CO2 cost splits with building efficiency rating, and all legal disclosures. Templates are versioned to accommodate future legal changes.
Implementation journey
Phase 1: Discovery and architecture design
- Deep-dive analysis of 95-page prompt and client's Gemini-based prototype
- Workflow mapping: property manager pathway vs. self-managed pathway
- Legal requirements documentation (BGB, HeizKV, CO2KostAufG)
- Architecture design: agentic orchestration, calculation separation, RAG integration
- Technology selection: AWS Bedrock (Claude 4.5 Sonnet) over client's Gemini approach
- Infrastructure planning: Terraform-based IaC, dual-repo structure (frontend/backend)
Phase 2: Core development and iteration
- Sprint 1: Document processing pipeline (OCR, classification, extraction, validation)
- Sprint 2: Conversational orchestration engine with intent detection and workflow routing
- Sprint 3: Calculation engines with triple-validation and legal compliance checks
- Sprint 4: PDF generation system and end-to-end testing with real landlord documents
- Agile sprints with weekly client demos
- Real-world document validation (handwritten invoices, poor-quality scans)
Phase 3: Launch and stabilization
- Beta launch with 50-100 early-access landlords
- Performance optimization and token cost monitoring
- User onboarding refinement based on actual usage patterns
- Comprehensive error logging and monitoring dashboard setup
- Final legal review with German rental law attorney
Phase 4: Growth and optimization
- Monitor peak-season performance (June-December billing cycle)
- Implement advanced features: year-over-year property data persistence, proactive legal update notifications, multi-agent workflows for complex scenarios
- A/B testing for UX optimization and conversion rate improvement
- Cost optimization: model fine-tuning, prompt compression, caching strategies
- Expansion to adjacent use cases: rent increase calculations, contract clause generation
Business impact and expected results
The agent is live at agent.reperks.de and open for anyone to try. Watch the Perks agent demo
What landlords achieve with Perks
Time back in your life
Complete annual billing in 15 minutes instead of 20+ hours. That's 98% of your time backUpload documents once and move on. No more hours retyping invoice dataSkip the legal research. Perks validates compliance automaticallyPeace of mind
Zero risk of mathematical errors that could cost thousands in disputesComplete legal protection. Every statement meets German rental law requirementsFull transparency. You see and approve everything before sending to tenantsBetter landlord experience
Natural conversation replaces confusing forms. Just talk to Perks like a colleagueWorkflow adapts to you, whether you have a property manager or handle everything yourselfScales effortlessly from 1 property to 100+
Strategic advantages gained
Market differentiation through AI-first design
While competitors force landlords into rigid form-based interfaces, Perks meets users in natural conversation. Landlords can upload documents in any order, ask questions mid-process, and receive intelligent guidance through legal complexities. This is a fundamentally different approach in a market dominated by legacy software.
Regulatory moat
Deep integration of German rental law (BGB, HeizKV, CO2KostAufG) creates defensibility. Competitors can't simply copy the conversational interface. They must replicate the entire legal knowledge layer and maintain it as regulations evolve. Perks' automated compliance validation becomes increasingly valuable as German climate legislation introduces new requirements (e.g., evolving CO2 cost allocation rules).
Platform foundation for expansion
The agentic architecture we built supports Reperks' broader vision beyond billing. The same conversational orchestration, document processing, and legal validation infrastructure can power adjacent use cases: rent increase calculations (§558 BGB), rental contract generation, tenant communication templates, and maintenance cost tracking. This positions Reperks to expand across the full property management lifecycle.
Data flywheel
Each completed billing adds structured property data (units, tenants, contracts, historical costs) to the database. Year-over-year, the system becomes smarter: pre-filling data, detecting anomalies by comparing to previous years, and requiring less user input. This creates switching costs and network effects as landlords build comprehensive property histories in the platform.
Architectural decisions enabling future scale
Multi-model flexibility
AWS Bedrock's model routing enables the client to optimize cost/performance as AI models evolve. Currently using Claude 4.5 Sonnet (Extended Thinking), they can switch to more cost-effective models for simple tasks or more powerful models for complex legal reasoning without code changes.Separation of concerns
By moving calculations into deterministic code (rather than relying on LLM arithmetic), we created a reliable foundation that can integrate third-party verification services or even blockchain-based audit trails as the market demands. The LLM handles only the aspects it excels at: conversation and reasoning.Observability and learning
Comprehensive logging of user interactions, document uploads, and calculation validations creates a rich dataset for continuous improvement. Future ML models can be trained on this data to improve document classification, extraction accuracy, and conversation quality without human labeling.
Key takeaways and lessons
What's making this successful:
Right-sized architecture for MVP stage
We're resisting over-engineering (e.g., multi-agent systems, vector databases for RAG) in favor of simpler patterns that meet current needs. This keeps the 6-week timeline achievable while designing for future evolution. Complex systems can always be added. Simplicity can't be retrofitted.Hybrid AI + deterministic approach
LLMs excel at reasoning but fail at precise arithmetic. We separated concerns aggressively. The LLM orchestrates conversation and explains legal requirements. Hardened code handles all calculations. This hybrid model is replicable across any regulated industry where compliance and precision are non-negotiable.Human-in-the-loop as feature, not bug
Rather than pursuing full automation (which would fail given document quality variability), we designed explicit confirmation flows that build user trust. Users see the system's work transparently and feel in control. This is especially critical for demographics unfamiliar with AI, where "magic" creates anxiety rather than delight.Legal knowledge as structured data
Instead of embedding laws as raw text in prompts (the client's initial approach), we structured legal requirements as JSON schemas with rule definitions. This makes the system testable, maintainable, and auditable. Essential for regulated domains. Future legal changes require updating schemas rather than rewriting prompts.
Technical innovations:
Context-aware calculation validation
Our triple-validation approach (percentage-based, unit-based, reverse-calculation) doesn't just catch errors. It provides context-specific error messages. When validation fails, the system can explain why in terms of German rental law (e.g., "Total allocation exceeds 100% because heating costs must follow HeizKV distribution rules"). This level of domain-specific error handling is rare in AI applications.
Document-type-specific extraction pipelines
Rather than generic OCR, we implemented classification-first processing that routes documents through specialized extraction templates. This pattern is replicable across any document-heavy automation: legal discovery, medical records processing, financial audits. The key insight: classification accuracy > extraction accuracy in determining overall system reliability.
Cost-aware model routing
We implemented basic but effective cost controls: expensive models (Extended Thinking) only for complex legal reasoning, cheaper models for simple extractions, aggressive caching for repeated legal queries. As LLM costs remain a key barrier to AI adoption, this engineering discipline must become standard practice.
HyperSense capabilities featured
AI Development
Agentic workflow architecture, AWS Bedrock integration, multi-model orchestration, cost-optimized LLM engineeringCustom Software Development
Rapid MVP methodology, user research integration, technical roadmap aligned with market windowsCloud Migration and Infrastructure
Serverless AWS design, scalability planning for 10K+ concurrent users, Infrastructure as CodeIntelligent Document Processing
OCR with confidence scoring, classification-first extraction, human-in-the-loop validation
Related case studies
Insurance & Healthcare Technology
Signal Iduna: Digital health insurance platform
Enterprise insurance platform serving 300,000+ policyholders with 2+ year engagement and 8-person team
Chemical transport
Trusted Carrier: Logistics ecosystem
Three-product digital ecosystem processing 40,000+ documents weekly across 13 countries
Recycling & Waste Management
Hamilton Perkins: AI chatbot for B2B/B2C
Dual-purpose conversational AI using AWS Bedrock and Anthropic Claude
Ready to build production-grade AI agents?
Building reliable AI agents requires more than prompt engineering. It demands architectural rigor, domain expertise, and engineering discipline to handle edge cases, cost controls, and regulatory requirements.
Whether you're tackling document automation, conversational AI, or compliance-heavy workflows, our team brings proven experience shipping production systems that users trust with high-stakes tasks.
Tell Us About Your Project
Describe what you're building, your timeline, your KPIs, and any technical constraints. We'll review your project and get back to you with a clear next step.Let's talk about your project