El Empleado Juan

Co-founder. A B2B SaaS for biometric attendance and automated payroll aimed at Latin American small and medium businesses.

Three capture layers

No customer has a single operational reality, so the system clocks people in three ways and reconciles them:

• A kiosk with in-house facial recognition — detection with MTCNN/RetinaFace, embeddings with FaceNet/ArcFace, anti-spoofing, deployed on low-cost hardware.
• Hikvision terminals, for plants that already had hardware installed.
• Mobile clock-in with GPS and geofences, for staff on the road.

The other half: the law

Biometrics only answers who clocked in and when. What the customer actually buys is what comes next: Colombian payroll computed against the labour reform — Law 2101 of 2021, the 42-hour week, auditable premiums computed minute by minute. Every day worked is settled under the law in force on that day, even when a pay period straddles a change in legislation.

Juanito, the AI assistant, closes the loop over WhatsApp: clocking in, checking schedules and hours, requesting time off.

Real scale

Anchor customer: a food logistics operation with 900 employees in Medellín.

It started as a WhatsApp bot on Twilio that turned into a platform. The biometrics I began on the Electoral Tribunal kiosks, now in the hands of small businesses.

The pipeline, from camera to punch

Embeddings come out normalised, so cosine similarity is a dot product and comparing against a whole site is one matrix-vector multiply:

$$\operatorname{sim}(u,v)=\frac{u\cdot v}{\lVert u\rVert\,\lVert v\rVert}=u\cdot v \qquad\text{with}\qquad \lVert u\rVert=\lVert v\rVert=1$$

# The index is per SITE, not global: comparing against the 900 vectors of
# one plant is not the same operation as against the customer's 40,000.
index = site.embeddings()                     # (n, 512) float32, L2-normalised

def identify(face: np.ndarray, threshold: float) -> Punch | None:
    if not is_live(face):         # anti-spoofing first: a photo has no
        return None               # business reaching the matcher

    v = arcface.embed(align(face))            # (512,)
    similarities = index @ v                  # cosine, since both are unit
    i = int(similarities.argmax())

    # The threshold is calibrated on THIS site's data. The lab value does
    # not survive the first shift: a cheap camera, backlight at 6 a.m. and
    # people in a hurry move the whole distribution.
    return Punch(site.employee(i)) if similarities[i] >= threshold else None

All of this runs on the kiosk itself. Uploading the frame to a server would be easier to write and would turn every network outage into a plant that cannot clock in.