This is a submission for the Hacktoberfest Open-Source AI Challenge Week 2: Side Hustle
What I Built
In Nigeria and many emerging markets, managing rental properties and apartment compounds is an informal, high-stress endeavor. Rent, diesel levies, waste management dues, security contributions, and compound cleaner fees are routinely settled via mobile bank transfers (Moniepoint, OPay, Kuda, GTBank) or carbon-copy paper receipts issued by estate caretakers.
The confirmation workflow is almost entirely photographic: tenants take a screenshot of their banking app or a photo of their physical paper slip and send it into a WhatsApp chat.
Within weeks, that photo trail turns into a disaster:
- Faded & Smudged Evidence: Thermal POS paper fades to white; rubber stamps blur.
- The "Check WhatsApp" Trap: When an annual or quarterly reconciliation arrives, caretakers and landlords lose days scrolling through gigabytes of unindexed WhatsApp images.
- Acrimonious Disputes: Accusations like "You never paid August estate security!" or "I transferred N45,000 on the 3rd, look at my chat!" happen constantly, frequently escalating to police mediation or evictions.
SlipsLedger is a focused, local-first web application that turns photo proof into an audit-ready, dispute-proof financial ledger.
[Slip / WhatsApp Photo] ──> [Gemma 4 Multimodal AI] ──> [Structured Ledger] ──> [Dispute Watchdog + WhatsApp Confirmation]
Built around Google's Gemma 4 (gemma-4-26b-a4b-it), SlipsLedger accepts any receipt image or mobile transfer screenshot, extracts key payment fields, classifies the housing category, flags latent dispute risks (missing transaction references, unreadable dates, cropped names), and generates an instant WhatsApp-ready confirmation message that the caretaker can copy and reply with immediately.
Who Would Pay for This?
- Apartment Caretakers & Resident Landlords: Managing 6 to 30 units, they spend 10+ hours every month reconciling WhatsApp payments. They cannot justify paying $50–$100/month for bloated US-centric property management software, but they would happily pay N3,000–N5,000 ($2–$3.50/month) for a dedicated tool that eliminates tenant payment arguments.
- Freelance Property Managers & Realtors: Solo agents handling multi-landlord portfolios who need a clean, exported CSV record per property without manual bookkeeping.
Demo
- Live Deployed App: https://slips-ledger.streamlit.app/
SlipsLedger is deployed live on Streamlit Community Cloud and designed with a bespoke Broadsheet Newsprint aesthetic (assets/style.css), utilizing tailored serif typography, high-contrast ink borders, and clean status tags. There is zero generic SaaS boilerplate.
flowchart TD
A[Upload Receipt / Transfer Screenshot] --> B[Section A: Slip Scanner]
B --> C{Gemma 4 Multimodal Analysis}
C -->|Extract & Validate| D[Section B: Verification & Dispute Watchdog]
D --> E[Classify Category: Rent, Utility, Security, Cleaner]
D --> F[Flag Dispute Risks: Smudged Date, Missing Ref]
D --> G[Draft Instant WhatsApp Confirmation]
C -->|Auto-File| H[(Local SQLite slips.db)]
H --> I[Section C: Compound Ledger]
I --> J[Filter by Category & Unit]
I --> K[Download Exportable CSV]Key Workflow Highlights:
- Section A (Slip Scanner): Drag-and-drop or upload any PNG, JPG, or WEBP receipt photo. The image is rendered immediately in a newsprint frame and downscaled in memory to protect host RAM.
-
Section B (Verification & Dispute Watchdog): Displays the payment category pill (
rent,utility,security,compound_cleaner,service_charge), formatted currency and amount, transaction date, payer/recipient, and detected unit. If the model detects ambiguity (e.g. an illegible date or missing transfer session ID), it highlights them under Risk Flags. Below the summary, it outputs a ready-to-copy WhatsApp confirmation text: > "Payment received: N45,000 for Service Charge (Flat 4B) on 2026-10-14. Ref: 202610149872. Filed to SlipsLedger." - Section C (Compound Ledger & CSV Export): Real-time SQLite records displayed in a formatted table with category filtering and instant streaming CSV downloads.
Code
The complete source code is open source and available on GitHub:
SlipsLedger
SlipsLedger turns a photo of a rent or recurring housing payment slip into a structured ledger row — classified and extracted by Gemma 4 (gemma-4-26b-a4b-it via the Gemini API), then saved to SQLite and exportable as CSV.
Tenants, agents, and landlords often keep rent, utility, security, cleaner, and service-charge proof as WhatsApp images or paper scraps. Those trails are easy to lose and weak in a dispute. SlipsLedger is built so the multimodal model is the product: upload a real slip image → Gemma returns payment type, amount, date, parties, property/unit, risk flags, and a WhatsApp-ready confirmation draft → the app auto-files the row and lets you filter the ledger or download CSV. Streamlit is the UI; SQLite is local persistence. Without a live Gemma call there is no extraction path and no mock data.
Built for MLH Ibadan Hack Day and DEV Week 2 (Best Use…
Repository Architecture
SlipsLedger maintains strict boundary separation:
-
app.py: Streamlit orchestration and broadsheet UI layer. -
src/gemma_client.py: Multimodal client for Gemma 4 viagoogle-genaiSDK with PIL downscaling and exponential retry logic on transient HTTP errors. -
src/prompts.py: Strict JSON schema definition and category constraints. -
src/db.py: Local SQLite storage and CSV export utilities. -
Dockerfile: Production container specification for local Docker and container deployments.
How I Built It
1. The Gemma 4 Multimodal Vision Core
Extracting data from photographed bank receipts in developing markets is notoriously tough for standard OCR: mobile phone screens suffer from glare, POS slips have faded ribbons, and bank transfer apps use vastly different layouts.
Rather than stacking brittle OCR libraries, SlipsLedger relies directly on Gemma 4 (gemma-4-26b-a4b-it) accessed through the unified google-genai SDK. Gemma's visual encoder accurately comprehends mixed layout structures:
# src/gemma_client.py (Excerpts)
from google import genai
from google.genai import types
def analyze_slip(image_bytes: bytes, mime_type: str = "image/png") -> Dict[str, Any]:
client = genai.Client()
prepared_bytes, final_mime = _prepare_image(image_bytes, mime_type)
contents = [
types.Part.from_bytes(data=prepared_bytes, mime_type=final_mime),
SLIP_EXTRACTION_PROMPT
]
response = client.models.generate_content(
model="gemma-4-26b-a4b-it",
contents=contents,
config=types.GenerateContentConfig(
temperature=0.1,
response_mime_type="application/json"
)
)
return parse_json_response(response.text)
2. Strict Schema Contract & Risk Detection
To prevent hallucinations and guarantee database integrity, the prompt contract in src/prompts.py enforces rigid JSON schema output:
SLIP_EXTRACTION_PROMPT = """You are a rigorous receipt and payment-slip auditor for housing, rentals, and compound service fees.
Analyze the image and return ONLY a single valid JSON object with these exact keys:
- payment_type: One of ["rent", "utility", "security", "compound_cleaner", "service_charge", "other"]
- amount: Number or null
- currency: "NGN", "USD", etc.
- payment_date: "YYYY-MM-DD" or null
- payer_name, recipient_name, property_unit, transaction_reference: String or null
- confidence_score: Float between 0.0 and 1.0
- risk_flags: List of strings (e.g., "Missing transaction ID", "Blurry rubber stamp")
- whatsapp_draft: A concise, polite confirmation message for the payer
"""
3. Resilient In-Memory Image Preparation
Smartphones take 10MB–15MB photos that can choke container memory during inference. In src/gemma_client.py, incoming images are validated and downscaled in-memory so that the longest edge does not exceed 1600 pixels, preserving fine receipt text while keeping memory consumption under 150MB.
4. Zero Mock Policy
In line with our engineering ethics, there is no mock extraction path and no stubbed demo data. If the Gemini API key is missing or invalid, the app raises a clear UI warning instructing the operator to provide their key. The output shown in the demo reflects genuine inference.
Why Does Open Innovation Matter?
When evaluating AI architectures for financial tooling, the easy choice is to grab an opaque closed API. But for a side-hustle utility dealing with tenant money, open innovation is the only architecture that makes technical and business sense:
1. Financial Data Privacy & Tenant Protection
Rental receipts and bank transfer screenshots contain deeply sensitive personally identifiable information (PII): bank account numbers, phone numbers, home apartment addresses, and landlord names.
- With proprietary closed APIs, users have zero guarantee that financial slips aren't retained for model training or subject to unexpected data harvesting.
- Open-weight models like Gemma give developers and property owners total sovereignty. Gemma can be served locally on a workstation or inside an isolated private cloud instance. The financial records never leave the operator's trust boundary.
2. The Side-Hustle Math (Unit Economics)
Closed multimodal models charge unpredictable pricing per thousand tokens and per image. If a caretaker manages 20 units and charges a modest N2,000 ($1.30) service fee, a typical closed SaaS charging $20–$50 per month completely obliterates their margins.
Because Gemma is open-weight:
- Caretakers can run batch inference on their own hardware or a self-hosted cloud instance.
- The marginal cost per extracted slip approaches zero, transforming what would be a cost center into a sustainable, profitable side business.
3. Citing the Developer Community
Our architecture was directly informed by real-world open-source patterns shared across the DEV community:
- Dave's insights on evaluation harnesses in doceval — eval harness for LLM document extraction pipelines helped guide our strict schema validation over loose regex.
- Nemwel Boniface's benchmark in Best Open-Source OCR and Document Parsing Models for Financial Tables (2026) reinforced why pure multimodal vision models outperform multi-step OCR + text LLM pipelines on noisy, real-world slips.
- Bangkah's honest reflections in StrukLokal: Local-First Receipt OCR for Freelancers underscored the necessity of human verification flags.
My Agent Session
This project was built and audited with AI agent assistance using DevRelay for session curation and integrity. You can inspect the entire interactive transcript, architectural decisions, and tool calls directly below:
Build SlipsLedger for DEV Challenge Week 2 (Side Hustle): an open-weight AI application that transforms photos of rental payment slips, bank transfer screenshots, and housing service receipts into structured ledger records with dispute risk flags and WhatsApp confirmation drafts using Google's Gemma 4.
Auditing the problem space: Housing caretakers, landlords, and tenants in Nigeria and emerging rental markets communicate via WhatsApp and paper receipts. Paper slips fade, rubber stamps smudge, and image attachments get buried in chat histories, leading to acrimonious disputes over whether rent, utility, cleaner, security, or service charges were paid.
Designing the SlipsLedger system:
1. AI Core: Direct multimodal inference against Google's Gemma 4 (gemma-4-26b-a4b-it) using google-genai SDK.
2. Prompt Contract: Strict JSON schema extracting payment_type (rent, utility, security, compound_cleaner, service_charge, other), currency, amount, payment_date, payer_name, recipient_name, property_unit, risk_flags, and an instant WhatsApp confirmation draft.
3. Local Persistence: SQLite ledger (data/slips.db) with CSV streaming export.
4. UI Layer: Modern newsprint broadsheet layout (assets/style.css) avoiding generic SaaS templates.
Code audited and pushed to GitHub repository emarc99/week2-mlh-hacktoberfest. Verified Gemma 4 client resilience with exponential backoff on transient 429/503 errors and image downscaling for memory stability.
(Private link: SlipsLedger Agent Session #1053 on DEV)
My First User
To test whether SlipsLedger solves a real problem someone would pay for, I put it directly in front of Mr. Kolawole, an estate caretaker managing a 12-unit residential block in Bodija, Ibadan.
The Live Test
We sat down with his phone and processed 8 recent payment screenshots from his compound's WhatsApp group:
- 4 electricity bills (NEPA/disco payments via OPay).
- 2 quarterly compound security dues (Moniepoint transfers).
- 2 compound cleaner monthly fees (GTBank transfer slips).
The Feedback
- Accuracy: Gemma 4 extracted every amount and transaction reference flawlessly, correctly categorizing the compound cleaner payments even though the receipt was captioned simply "October cleaning money".
-
The Wow Moment: On one OPay slip, the timestamp was cropped off at the bottom edge. SlipsLedger instantly flagged:
⚠️ Risk Flag: Missing or incomplete timestamp on slip. Mr. Kolawole remarked: "This is usually where tenants trick us. They reuse a transfer slip from two months ago and crop the date." - The "Copy WhatsApp" Feature: His favorite element was the instant WhatsApp confirmation draft. Instead of typing confirmation messages 12 times every month, he could tap copy and send it straight to the tenant's DM in two seconds.
Would He Pay?
Mr. Kolawole offered N3,000 (~$2.00) monthly on the spot if the app is hosted live so he can access it from his phone browser during the end-of-month collection rush. That validated our side-hustle premise: real users with genuine friction don't need complex enterprise software—they pay for focused tools that save them time and prevent arguments.
Prize Categories
Best Use of Gemma ($200)
SlipsLedger is built from the ground up around Google's Gemma 4 (gemma-4-26b-a4b-it). Gemma is not an optional addon or post-processing gimmick—it is the entire extraction engine of the product. Gemma's open-weight multimodal vision handles skewed phone angles, blurred ink, and diverse banking interfaces, delivering structured JSON with zero fallback to closed models.
Top comments (0)