AI for physical process transformation

Cobalt measures how work happens on a manufacturing floor, using camera video alongside the plant's existing data. It shows where the time goes, what to change, what each change is worth, and whether the change held.

How work actually happens is rarely written down. Most plants run on what people know, and decisions about staffing, layout and automation get made without a measured baseline.

From baseline to tracked result

  1. 01

    Baseline

    Video from the floor, joined to existing plant data such as MES and ERP records, is converted into measures for each station: the share of the shift spent on value-added work (the work that changes the part), how long a unit takes and how much that varies, units finished per day against plan, and time spent on rework. In manual welding, arc time typically runs 15 to 30 percent of the shift.

    Time use by station

    Share of shift, ten shifts pooled

    Cams 1 to 4 · 4 stations
    Fit-up Station A0% value-added
    39%
    29%
    13%
    13%
    Weld Booth 30% value-added
    18%
    38%
    17%
    11%
    16%
    Weld Booth 50% value-added
    26%
    37%
    14%
    9%
    14%
    SAW long seam0% value-added
    31%
    36%
    12%
    15%
    • Value-added work
    • Support work
    • Waiting
    • Walking
    • Breaks and idle
  2. 02

    Root cause

    Lost time is broken down by cause: waiting for parts, for the crane, for an inspector, or walking to the tool crib. The system identifies the bottleneck, the one station that limits the output of the whole line.

    Idle by cause

    Fit-up Station A, minutes per shift

    Fit-up Station A · Cam 1
    Waiting for the crane
    0
    Walking to the tool crib
    0
    Waiting for an inspector or a drawing answer
    0
    Waiting for consumables
    0
    Load by station, scheduled hours as a share of available
    62% Cut and form
    58% Roll
    105% Fit-up
    81% Weld
    94% NDE
    66% Finish
  3. 03

    Propose changes

    The plant receives a ranked list of specific changes. Each item links to the video behind it, by camera and time, so supervisors can check it.

    Ranked changes

    Estimated value in canisters per week

    Cams 1 to 4 · 10 shifts
    1. 1

      Stage kits and consumables with one material handler

      Fit-up, weld boothsCam 1 · 07:12Cam 3 · 09:41
      +0.4 / week
    2. 2

      Book the hold-point inspector when fit-up starts, not when welding ends

      Fit-upCam 1 · 10:26
      +0.3 / week
    3. 3

      Check the fit-up gap before the assembly leaves the station

      Fit-upCam 1 · 13:05Cam 4 · 13:48
      +0.2 / week
    4. 4

      Adopt the best-observed circ seam sequence at all four booths

      Weld boothsCam 3 · 08:15Cam 4 · 08:22
      +0.2 / week

    Each item links to the video behind it, by camera and time.

  4. 04

    Quantify

    A working model of the floor estimates what each change is worth in units per day and cost per unit. The same model is used to test larger decisions, such as a second shift or a welding robot, against the measured baseline.

    Throughput model

    With changes 1 to 5, against the baseline

    canisters per week
    3.3 → 3.3
    Constraint
    Fit-up → NDE
    Cost per unit
    $21,400 → $21,400
    Cut and form
    78%
    Roll
    73%
    Fit-up
    92%
    Weld
    96%
    NDE
    100%
    Finish
    83%

    Load: scheduled hours as a share of available hours. Relieving fit-up moves the constraint to NDE.

  5. 05

    Track

    Measurement continues after a change is made. The core measures update continually, so the plant can see whether the change was adopted and whether the number moved.

    Week-over-week time study

    Fit-up Station A, share of shift by week

    Fit-up Station A · Cam 1 · 8 weeks
    1
    2
    3
    4
    5
    6
    7
    8
    Change made after week 4
    Material handler starts staging kits
    Value-added work
    39% → 0%
See each step

One week to set up. Results start in week two.