No result found
This report needs a scored result to generate.
This is a static sample and does not depend on a scored result.
Back to the course page
PRAXORALAB Illustrative example · fictional company, for format reference only
AI & Robotics Feasibility Checklist
Tanjong Bay Marine Fittings Pte Ltd
Overall Feasibility Score Developing
A mixed profile: enough volume and a detectable defect pattern to justify a bounded pilot, held back by two absent foundations, a labelled image library and a named process owner, neither of which exists yet.
| Dimension | Score | Status | |
|---|---|---|---|
| Process stability | 2 / 3 | Developing | |
| Data availability | 1 / 3 | Blocker | |
| Failure-mode detectability | 2 / 3 | Developing | |
| Volume & throughput | 3 / 3 | On track | |
| Ownership & integration | 1 / 3 | Blocker |
Priority: Data availability
No labelled defect image library exists today. Rejected castings are pulled off the polishing line by feel and eye and dropped in a scrap bin, not photographed, dated or categorised by defect type, so a vision model has nothing to train against yet.
A nine-out-of-fifteen score is almost exactly where I expect a first-time vision-QC candidate to land, and it's worth being precise about which of these five numbers actually stops the pilot and which ones just describe work still to be done. Volume and throughput scoring a 3 tells me the business case is genuinely there: a single steady shift is more than enough daily volume to make a fixed camera worth the capex, and Tanjong Bay's process owner already tracks output by product code, which is more operational discipline than most first-time candidates bring into the room. Process stability at a 2 is also a real asset, not a technicality: porosity, flash and incomplete fill are exactly the kind of stable, repeatable visual pattern a vision model can learn, which is the same starting point the pharmaceutical vision-QC line I commissioned had before it ever went near a conveyor.
The two 1s are where I would stop the conversation, and I would stop it on data availability before I stopped it on ownership, even though both score the same. The pharmaceutical line I built ran on roughly six thousand labelled sample images before it went anywhere near a production camera; Tanjong Bay currently has zero, because rejected castings go into a scrap bin unphotographed and uncategorised. That is not a documentation gap you close in a week. It is months of deliberately keeping and labelling rejects before a model has anything to train against, and it should start on day one of any pilot regardless of what else gets decided, because nothing downstream works without it.
Ownership scoring a 1 is a different kind of problem, and an easier one. Two line leads alternating an informal visual check is a coordination gap, not a technical one, and naming one accountable owner for the polishing line's QC step is a conversation, not a data-collection programme. I would close that gap in the first pilot-scoping meeting, not somewhere in the first ninety days.
What I would flag as genuinely underrated on this page is failure-mode detectability's own caveat. Hairline stress cracks are the defect category most likely to drive an actual warranty claim, and no camera system, however well trained on porosity and flash, catches them. Any pilot scope that doesn't explicitly carve that failure mode out, and route it to a separate ultrasonic or dye-penetrant check, is quietly promising more coverage than the technology can deliver, and that gap is exactly what the next page's scope document has to name honestly rather than let a vendor's demo paper over.
PRAXORALAB Draft Pilot Scope
Built for one real process brought into the room. Tanjong Bay Marine Fittings Pte Ltd's example use case: Machine-vision quality control for cast marine deck-hardware fittings, ahead of the polishing line. The scope below names what is in, what is data-dependent, and what is deliberately excluded, not left as a general ambition to "add AI".
In scope — Vision QC station, polishing-line infeed
One camera station mounted at the polishing line's infeed conveyor, inspecting each cast fitting before it enters the buffing wheel.
Three defect classes only for the pilot: surface porosity, incomplete fill, and flash at the parting line.
A 90-day trial on the single existing shift, not a second shift or a second product line.
Data & training set
Build a labelled image library from rejects the polishing line already pulls by hand, starting week one, targeting at least 2,000 labelled images across the three defect classes before model training begins.
An external QA pass on the label set before training, since a labelling error is the most common cause of a vision model's early false-reject rate.
Failure modes to detect, and the one deliberately excluded
Porosity, incomplete fill and flash are all visually detectable at standard camera resolution and are the pilot's in-scope defect classes.
Hairline stress cracks, the defect class most linked to warranty claims, are explicitly out of scope for vision and routed to a separate dye-penetrant spot check, not silently absorbed into the vision system's claimed coverage.
Integration points
Camera and reject-diverter arm retrofit onto the existing infeed conveyor; no new conveyor purchase.
Reject log routes into the daily production spreadsheet the line already uses, not a new dashboard requiring separate reporting discipline.
Explicitly out of scope for this pilot
Line-speed increases, a second product line, and any downstream automated sorting are excluded from the 90-day pilot and reserved for a second-phase business case if the pilot's defect-escape numbers justify it.
The scope on the opposite page is deliberately narrow, and narrower than most first-time pilots want to start, so it's worth explaining why each boundary is where it is rather than somewhere wider. Restricting the pilot to three defect classes, porosity, incomplete fill and flash, is not a technology limitation, it's a sequencing decision: those three are the defects the polishing line's own rejects already show a stable, repeatable visual signature for, which is exactly what a first model needs to learn cleanly before anyone asks it to handle a harder or rarer defect type. I have watched pilots try to cover every defect class on day one and end up with a model that is mediocre at all of them rather than genuinely reliable at three.
The hairline-stress-crack exclusion is the line I would defend hardest in a real scoping session. It would be easy, and tempting, to let a vendor's proposal quietly imply the vision system "catches quality issues" broadly, without naming what it cannot see. I would rather the pilot scope say plainly that stress cracks need a separate dye-penetrant check, because the alternative, an implied but unstated gap, is how a warranty claim slips through eighteen months from now and nobody can explain why the "AI quality system" didn't catch it.
The integration choice, retrofitting the camera and diverter arm onto the existing infeed conveyor rather than specifying a new one, is a cost and timeline decision as much as a technical one. A new conveyor purchase adds procurement lead time and a second capex line the EDG file would need to separately justify; a retrofit keeps the pilot's cost, and its risk, bounded to the one system actually being tested. Routing the reject log into the daily production spreadsheet the line already uses, instead of a new dashboard, is the same instinct: don't ask the floor to adopt two new habits, the camera and a new reporting tool, in the same ninety days.
The out-of-scope line at the bottom is the one I would insist stays in every scope document I sign off on, not just this one. Naming what a pilot is not doing, a second shift, a second product line, automated downstream sorting, protects the pilot's own numbers from being judged against a scope it was never given the budget or time to meet.
Tanjong Bay Marine Fittings Pte Ltd and its scope details are invented for this sample only, to show the shape of the output, not a real client's actual pilot. In the session, this scope is built from the process a participant brings, not assigned from a template.
PRAXORALAB Enterprise Development Grant Readiness Note
The four questions a grant-aware reviewer will actually ask, answered for Tanjong Bay Marine Fittings Pte Ltd's example use case above, not left as an abstract eligibility summary.
Reducing the defect-escape rate past the polishing line, measured against a manual baseline defect-escape rate recorded over the pilot's first two weeks before the camera goes live, not an assumed improvement.
Enterprise Development Grant support typically co-funds a defined percentage of qualifying costs against at least two comparative vendor quotations. Tanjong Bay has one vendor quotation drafted and needs a second before an application can be filed at all.
At least one internal operator trained to add newly labelled rejects to the training set and trigger a retrain, so the model doesn't quietly go stale the first time the alloy supplier changes and nobody who understands it is available.
Dated before-and-after defect-escape figures, the vendor's signed project-completion sign-off, and the original comparative quotations, retained together, not scattered across a vendor folder and an email thread.
Read together, these four answers are the difference between a grant application that survives Enterprise Singapore's own review and one that gets sent back for more information, and I would score Tanjong Bay's draft as close but not yet submittable. Question one's answer, measuring against a manual baseline recorded before the camera goes live, is the right instinct and the one first-time applicants most often skip. I have seen grant-funded projects claim an improvement in defect rate with no baseline behind the claim at all, which is not a technicality a reviewer waves through, it is the difference between a measured result and an assertion.
Question two's answer is the one I would push hardest on before this file goes anywhere. One vendor quotation is not a grant application; Enterprise Singapore's own comparative-quotation requirement exists specifically to stop a business case being built around whichever vendor got in the room first. This is not bureaucracy for its own sake. On the Raymang Eggs robotics line, an EDG-funded project I delivered end to end, the second quotation surfaced a materially different integration approach that changed the final scope for the better; treating the second quotation as a box-ticking formality would have meant missing that.
Question three's answer, naming an internal operator who can retrain the model rather than leaving that entirely to the vendor, is the single most consequential line on this page, and the one I would flag as underweighted if I only had time to read one answer. A vision system that only the vendor can retrain is not a capability Tanjong Bay owns, it is a subscription with an on-site camera attached, and the first time the alloy supplier changes and the model's accuracy quietly drops, nobody internal will know why. A grant is meant to fund capability transfer, not just capital equipment, and this answer is where that actually shows up or doesn't.
Question four's documentation list is complete on paper but untested until someone actually has to produce it under audit. My honest read of this specific business: dated before-and-after figures and a signed vendor completion sign-off are achievable; the discipline to keep the two original quotations filed together with them, rather than buried in an inbox eight months later, is the item I would personally check on again at the sixty-day mark, not assume is handled because it's written down here.
Tanjong Bay Marine Fittings Pte Ltd's readiness note above is invented for this sample only, to show the shape of the output, not a real applicant's actual grant file. Grant eligibility and co-funding terms should always be confirmed directly with Enterprise Singapore, not inferred from this illustration.
PRAXORALAB 90-Day Pilot Roadmap
The feasibility checklist, pilot scope and EDG readiness note above, brought together into one sequenced plan, ordered by what each phase's output feeds into next rather than by department.
- Name one accountable owner for the polishing line's QC step, replacing the informal two-lead handover.
- Start pulling and labelling scrap-bin rejects into the three defect classes, targeting 500 labelled images by day 30.
- Record a two-week manual defect-escape baseline before any camera is installed.
- Get the second comparative vendor quotation drafted for the EDG file.
- Continue building the labelled image library toward the 2,000-image target.
- Install the camera and reject-diverter arm on the infeed conveyor during a scheduled maintenance window.
- Route the reject log into the existing daily production spreadsheet.
- Submit the EDG application with both quotations and the recorded baseline attached.
- Run the vision system live for the pilot's single shift, comparing its defect-escape rate against the pre-camera baseline.
- Train the named internal operator to add new rejects and trigger a retrain.
- Present the pilot's defect-escape numbers and EDG documentation for the next budget conversation.
The reason this plan runs foundation, then build and data, then trial and sign-off, and not some faster-looking order, is that the first thirty days produce two things nothing later in the plan can substitute for: a recorded baseline and a genuinely accountable owner. Skip straight to installing the camera in week two, which is the instinct I see most often from an operations team eager to show progress, and there is no pre-camera defect-escape figure to measure the pilot's actual improvement against, which means the ninety-day result at the end is a number with no comparison point, which is not a result Enterprise Singapore, or Tanjong Bay's own board, should accept as evidence of anything.
Days 31 to 60 are where I would watch most closely for the plan quietly slipping, because installing hardware during a scheduled maintenance window is a hard deadline with a visible outcome, while continuing to build the labelled image library toward two thousand images is unglamorous, ongoing work that produces no single satisfying milestone. It is exactly the kind of task that gets deprioritised the moment the camera physically arrives and everyone's attention shifts to watching it work. I would name a specific person accountable for the labelling target hitting two thousand by day 60, separately from whoever manages the camera installation, precisely because those two workstreams compete for the same people's attention under normal operational pressure.
Days 61 to 90 are where the pilot either earns its next budget conversation or doesn't, and the two items I would not let get compressed under time pressure are training the internal operator and presenting the actual defect-escape comparison, not a qualitative impression of how the trial went. "The camera seems to be catching things" is not the same claim as a measured defect-escape rate compared against the recorded baseline, and only the second one survives a sceptical question from whoever signs off on the next phase's budget.
What "done" looks like at day ninety is not a working camera. It is a working camera, a trained internal operator who doesn't need the vendor on the phone to retrain it, and a measured, baselined defect-escape number Tanjong Bay can put in front of its next budget conversation without a caveat. That is a materially more defensible position than the one most ninety-day pilots actually reach.
Every score, quote and figure on these four pages is invented for Tanjong Bay Marine Fittings Pte Ltd, a fictional company, so the format of what a participant leaves with can be judged before enquiring. It is not a real client's deliverable, and no organisation named Tanjong Bay Marine Fittings Pte Ltd is a Praxora Lab client. The session itself assesses your own organisation's own process, in the room, on the day.
3.5 hours, one session, facilitated by Veronica Loh.