How LBS Tech Eliminated Fake Delivery Sign‑Offs at Daily Fresh

Background

In November 2014, Daily Fresh entered the fresh‑food e‑commerce market with an innovative “front‑warehouse” model, aiming to provide abundant, high‑quality fresh products and ultra‑fast home delivery. The model quickly gained market recognition, positive consumer feedback, and continuous investor support. Within three years, Daily Fresh achieved profitability in first‑tier cities and grew into an industry unicorn.

Daily Fresh heavily relies on LBS technology. Users of the Daily Fresh app or mini‑program first obtain their location or the most recent delivery address coordinates, then check whether the coordinate falls within the electronic fence of a front‑warehouse. Once a warehouse is identified, product browsing, ordering, and subsequent fulfillment are handled by that warehouse. During fulfillment, orders are assigned based on the user’s location, routes are planned from the warehouse to the user, couriers follow the route, and when a courier signs for delivery, the courier’s location determines whether the sign‑off status can be changed. Thus, LBS is the foundation of Daily Fresh’s business operation.

Over the past year, the Daily Fresh LBS team started from the address database, continuously strengthening the LBS technical system.

Address Database Construction

What is an address database?

An address database is a collection of delivery addresses derived from national standard place names, combined with logistics characteristics, and organized hierarchically. It includes administrative divisions, streets, house numbers (communities, buildings), unit numbers, postal codes, segment numbers, and basic property types, but excludes enterprise or personal information. The address database is a critical foundational data set for instant‑delivery logistics systems.

Self‑built address database

In the Daily Fresh app or mini‑program, users must create a delivery address before ordering. Creating an address requires selecting a POI (Point Of Interest). Incorrect or missing POIs prevent order placement. Before building its own address database, Daily Fresh relied on third‑party map services for nearby location search and recommendation. Those services provide generic scenarios that are not well suited for instant fresh‑food delivery, and address accuracy depends entirely on the map provider. Manual verification showed that in a certain community with 26 POIs, 3 were erroneous (11.5% error rate). Among 2,670 delivery addresses generated in that community, 54 were “over‑distance” (beyond 300 m) and 17 were severely over‑distance (beyond 1 km), making them undeliverable. Therefore, Daily Fresh decided to build its own address database optimized for the instant‑delivery scenario.

First, a confidence model based on delivery‑sign‑off distance was established, accumulating a large number of high‑precision POIs. Inaccurate POIs discovered later were corrected and saved back to the database. In the address‑search UI, POI recall rates were improved, non‑standard address weights reduced, detailed building POIs were prioritized, and ambiguous community POIs were ranked lower. Front‑end guidance and back‑end enforcement ensured standardization. The self‑built address database was completed in Q3 2020, with primary POI sources from the internal database and supplemental POIs from third‑party services. The database now supports address management, courier dispatch, sign‑off interception, and fake‑sign‑off detection, as shown in the architecture diagram below:

LBS Technical System Construction

During long‑term iteration of the delivery business, Daily Fresh accumulated various LBS capabilities. After the address database was built, LBS capabilities expanded along point, line, and surface dimensions, deepening existing functions and supporting fake‑sign‑off interception, community penetration rate calculation, and more. The following diagram shows the LBS capability panorama at the beginning of 2021:

Especially in fake‑sign‑off governance, the LBS system played a crucial role.

Fake Sign‑Off Governance

In online shopping, users sometimes see a package marked as “signed” before actually receiving it. This “fake sign‑off” occurs when a parcel is marked as delivered without being handed to the user, a common issue in the courier industry. In instant delivery, where service windows are typically 30–60 minutes, couriers facing potential penalties for late delivery may pre‑emptively sign off to avoid penalties. Without robust detection and appropriate penalties, more couriers exploit this loophole for personal gain.

Fake sign‑offs severely damage user experience, generate complaints, and cause user churn. From a management perspective, they distort delivery‑completion metrics, affecting decision‑making.

Daily Fresh’s fake‑sign‑off complaint rate remained high; many users complained daily, and most of those users eventually churned. Considering the high acquisition cost, fake sign‑offs were identified as the biggest delivery pain point in Q3 2020.

Challenges

The ideal solution would be immediate user sign‑off upon receipt, but in instant delivery most users do not click sign‑off right away. A common workaround is distance interception: at sign‑off, the courier’s real‑time GPS coordinates are compared to the delivery address coordinates; if the distance is within a threshold (e.g., 300 m), sign‑off succeeds, otherwise it is blocked. Daily Fresh initially used this method, but practical issues emerged, grouped into several categories:

Inaccurate Delivery‑Address Coordinates

Map providers’ POIs are sometimes inaccurate. While residential POIs are relatively accurate, complex or high‑security communities have lower accuracy due to manual collection difficulties.

Some users select a POI that represents an entire community or campus rather than a specific building. For example, a large community in Guangzhou has its POI centered, while the outskirts can be over 1.2 km away, exceeding the 300 m sign‑off radius.

User input errors: 1–2 % of users enter mismatched house numbers, causing significant coordinate gaps between the chosen POI and the actual location.

These inaccuracies cause distance calculations to deviate from reality, sometimes preventing couriers from signing off even when standing at the doorstep.

Location Drift

Location is obtained via satellite (GPS, Galileo, GLONASS, BeiDou), base‑station, or Wi‑Fi methods. Satellite accuracy is 5–10 m, base‑station 10–2000 m, Wi‑Fi 10–200 m. Apart from software bugs, drift mainly stems from:

Environmental factors: tall buildings or bridges block satellite signals; indoor or elevator environments rely on weaker Wi‑Fi or base‑station signals.

Device limitations: low‑end courier phones produce poorer accuracy, leading to drift.

Both cases cause courier coordinates to shift, affecting distance calculations.

Restricted Delivery Areas

Certain zones such as military bases, high‑security residential compounds, and some campuses prohibit courier entry, making distance interception ineffective.

Offline SOP

Not every sign‑off beyond the distance threshold is fraudulent; couriers sometimes simply forget to sign after completing delivery.

Appeal Review Cost

Given scenario complexity, automated decisions inevitably produce errors, requiring a courier appeal channel and manual review. Two key concerns are:

1) Determination accuracy – low accuracy leads to excessive manual workload and potential abuse.

2) Efficient auxiliary tools – reviewers need sufficient information to make quick, accurate judgments.

Distance interception caused about 5 % of orders to fail sign‑off, and reviewers lacked enough data, leading to the method being discontinued after less than a week.

Results

After governance, the proportion of orders with sign‑off distances over 500 m changed as follows:

The share of orders with fake‑sign‑off complaints also shifted:

Data shows a substantial improvement, positioning Daily Fresh’s fake‑sign‑off governance at an industry‑leading level.

Technical Solution

The challenge analysis revealed that the problem exceeds typical business requirements and lacks public solutions, necessitating a technology‑driven approach.

Relying solely on distance interception is insufficient. The comprehensive solution adds abnormal‑sign‑off reporting. For reported orders, we aggregate courier information (waves, trajectories, delivery addresses) and POI data (community, building, POI confidence) to design an algorithm that automatically determines whether the courier actually visited the delivery address after signing.

The main workflow is illustrated below:

Location Correction

To address drift, we apply real‑time trajectory correction using a dead‑reckoning algorithm, which estimates the current position based on known locations, speed, and heading. While error accumulates over time, drift typically occurs only after entering buildings, where intervals are short, keeping error within acceptable bounds (≤200 m for 76.3 % of cases).

Trajectory Determination

Trajectory Denoising

Drift creates noisy points; we remove them to obtain a realistic trajectory.

By quantifying movement features (speed, time to next point, distance, turning angle), each point receives a quality score. Points below a threshold are classified as noise.

Location source and precision metadata from the positioning API also serve as denoising references.

Trajectory Quality Assessment

Low‑quality trajectories risk misjudgment, so we evaluate quality and set a minimum standard. Trajectories below the standard are excluded from automatic determination.

Assessment considers missing points and continuous gaps, which arise from transmission loss or filtered noisy points.

Stay‑Point Mining

A stay‑point is defined as a cluster of consecutive points where the time span < T and the pairwise distance < L, with T and L being system‑defined thresholds.

For building‑level POIs, a confidence model based on historical sign‑off data determines a maximum sign‑off distance; stay‑points within this radius are considered sign‑off stay‑points. For community‑level POIs, if a stay‑point falls inside the electronic fence of a community, university, or factory, it is also treated as a sign‑off stay‑point.

These capabilities—location correction, trajectory denoising, quality assessment, and stay‑point mining—constitute the core of the LBS system.

Trajectory‑Based Fake Sign‑Off Determination

Based on trajectory quality, we locate sign‑off stay‑points. If a stay‑point satisfying the criteria exists before the recorded sign‑off time, the order is deemed a legitimate post‑delivery sign‑off; if such a stay‑point appears only after sign‑off, the order is flagged as a fake sign‑off.

In production, the fake‑sign‑off detection accuracy reached 96.1 % (accuracy calculated as (total reported orders – (mis‑detections + missed detections)) / total reported orders).

Appeal Tool

Delivery Replay Visualization

Automatic decisions may occasionally misclassify. Couriers can appeal, and the appeal is reviewed manually using a delivery‑replay visualization tool that mirrors the automatic logic. The tool’s interface is shown below:

Technology‑Driven Process

The fake‑sign‑off project was initiated and led by the technology team, following these steps:

Analyze the problem and establish metrics (abnormal report volume, fake‑sign‑off count, accuracy, false‑positive/negative rates, and various trajectory issue statistics).

Design and build the end‑to‑end workflow (sign‑off interception, abnormal reporting, fake‑sign‑off determination).

Implement automated reporting dashboards to drive cross‑team collaboration.

Launch gray‑scale validation, analyze bad cases, and iterate improvements; incorporate manual judgments into test samples for future self‑testing.

After each workflow change, back‑test with samples to verify impact, aiming for ≥95 % detection accuracy.

Coordinate product and business management actions.

Through relentless technical effort, automated fake‑sign‑off detection achieved >95 % accuracy, enabling the business to impose economic penalties on couriers. After the penalties were communicated, courier behavior improved dramatically, and the fake‑sign‑off rate dropped sharply. Analysis confirmed a strong correlation between detection accuracy and complaint reduction, confirming the success of the technology‑driven governance.

Conclusion

Leveraging years of accumulated LBS capabilities, Daily Fresh’s technology‑driven approach substantially mitigated fake sign‑offs. Future work will further refine detection accuracy and extend LBS applications to additional business scenarios.