Most software built for factories has the same flaw. It is excellent at telling you what already happened and close to useless at telling you what to do next. Takorin is my attempt to design the opposite of that, and this is the story of how the research forced me to throw out the obvious version of the product before I built it.
Before I drew a single screen, I sat down with 3 plant directors, each running operations at one of the top four food manufacturers in the world. I went in expecting a feature wishlist. What I got was closer to a warning.
Not one of them opened the dashboard they already owned. One said it plainly: it tells me what happened, not what to do about it. (If you have ever bought enterprise software and then watched the team quietly go back to a spreadsheet, you know the feeling.) These are people running plants with thousands of staff and razor margins, and the most expensive system on their desk was the one they trusted least.
That is the problem I was actually being asked to solve.
Food manufacturing lives and dies by OEE. A few points on a single line can swing profitability, and even one point is often worth six figures a year. So when a shift goes sideways, it is not a reporting issue. It is a financial one.
The problem is that bad shifts rarely come from one obvious failure. They are the results of small signals compounding over time: a checklist completed late, a sensor drifting warm, a qualification gap on the line, a supplier document that was never updated. By the time a supervisor sees the pattern, the moment to prevent it has usually already passed.
Most plants already have the data to catch this earlier but they live in separate systems, forcing plant leaders to connect the dots manually instead of seeing risk in one place.
That gap is becoming more expensive. A false alert wastes a few minutes. A missed signal can cost a batch, disrupt production, or trigger a recall. And with FSMA 204 traceability requirements taking effect in 2028, many plants are also facing growing pressure from major buyers to prove they can catch and trace issues faster.
The product direction came out of a few consistent signals in the research. WEF, BCG, and IBM’s writing on agentic AI helped shape the core model: a system that can spot issues, decide what matters most, take the right next step, and improve over time. Reporting from Food Industry Executive pushed the concept beyond monitoring and toward action, reframing the product as something that helps plant teams respond to problems, not just watch them happen.
I spoke with 3 plant directors who gave some insights:
M
Every director already had a Manufacturing Execution System, and none of them described it as the thing they open first.
Fear
The anxiety was "what if it's confidently wrong, I act on it, and the shift fails anyway." Every one of them had been burned before.
Timing
They all described a supervisor walking into a shift and by the time they have the full picture, the decision moment is gone.
Takorin sits on top of the systems plants already use. Each category does its job well, but each stops short of connecting signals across production, quality, maintenance, and suppliers. Production systems are strong at tracking what happened on the line. Quality systems are built for audits and corrective actions. Supplier and traceability systems manage certifications and lot tracking.
No system owns the decision layer. Takorin is built to sit across these categories and turn fragmented signals into a single ranked decision, with clear context, traceability, and a human approval step. It also exposes data readiness as a visible score and evaluates quality, supplier, and operational signals together in one place.
Every team on the project wanted the same thing, which was to surface more of their data. More quality charts. More supply numbers. I came very close to building a beautifully thorough dashboard that answered every question except the one a director asks at 6:38am, which is just: what do I deal with first?
So I did the opposite of what the brief implied and started taking things off the screen. The front door became one ranked queue. Every signal shows up with a priority, an owner attached, and a clock running, all of it built from data the plant already had. I made it a standalone screen and the default landing view rather than a panel on the old home screen, because a panel competes with everything around it while a screen with nothing else on it makes the ranked list the only thing in the room.
Plants run human-only crews, robotic lines, or a hybrid. By the 3rd screen that needed to know which, I stopped writing the same conditional over and over and moved that state into one shared place that every screen reads. It cost more on the first feature than an inline check would have. It paid back immediately, because the next two screens plugged into the existing value without touching the original logic, and I later folded both of them into the shift view for the same reason. Good architecture is mostly the boring decision to not repeat yourself.
Shift, the intervention engine
This is the screen the whole product is organized around, and it exists because of the interview findings. It reads the risk signals at the start of a shift, ranks the findings by how much they will cost, and attaches two things a generic alert never has:
1. A precedent from the plant's own past, because a director trusts a recommendation grounded in their own history
2. A countdown to the moment the window closes, because a finding without a deadline is just trivia.
The trust stack: Data Readiness and visible confidence.
This is the direct answer to the fear every director named. A strip at the top of every screen shows the health of the data sources feeding the system, so nobody reads a recommendation without knowing how fresh the inputs are.
Data Readiness is about the platform itself rather than the plant, and it scores how much of what Takorin says can be trusted. More to the point, it names the specific reasons confidence is low, the same ingredient sitting under three different names across the MES, the ERP, and the supplier portal, and it gives an honest effort estimate to fix each one. Most factory AI hides its uncertainty. This product points straight at it, because the directors told me that hiding it is precisely how trust dies.
Agent Control, where the human stays in the loop.
Takorin behaves like an agent. Named agents watch their corner of the operation and propose actions, each carrying its reasoning, its target, and a confidence score. A person approves or overrides, and nothing consequential happens without that approval. The confidence score is tuned per agent based on how costly a wrong call would be, which is why a recommendation about a compliance filing carries a higher bar than one about scheduling.
CAPA, where corrective actions actually correct.
Corrective action management in this industry is mostly a documentation exercise. I built the CAPA Engine to change what closed means. An evidence gate makes a case impossible to close without proof attached. Pattern analytics surface when several separate cases share one root cause, which is the thing a manual register never catches. And a benchmark layer shows a QA director her closure rate as a percentile against comparable plants, because the research said competitive context moves this buyer more than an absolute number ever will. It exists to make an FDA inspection a non-event rather than a quarterly emergency.
Suppliers, the edge of the plant.
Every other module reads data generated inside the building. This one handles the risk that walks in from outside. It tracks certificate status against the production schedule, watches delivery ETAs, and runs a shelf-life optimizer that flags a lot about to expire before the run that needs it, while there is still time to reorder. It also carries a supplier compliance scorecard, which exists for two reasons. FSMA 204 is coming, and a score a supplier can see changes that supplier's behavior without anyone having to chase them.
Knowledge, against the retirement cliff.
The master operators who hold these plants together are retiring, and their judgment leaves with them. The Knowledge tab captures that tacit expertise and rates the risk of losing it. The most expensive thing a plant can lose is the person who knows why the oven runs warm on the gluten-free line, and nobody had a place to put that knowledge before it walked out the door.
Impact, the proof.
This screen closes the loop because the whole business case depends on being able to say that a Takorin recommendation led to a real result rather than a coincidence. Without this loop, the impact claims are just telemetry with a narrative stapled on.
I directed a staged AI pipeline rather than freehanding screens one at a time. There was a stage that audited the existing system for the real gap, a stage that pressure-tested the approach before any pixels got drawn, a build pass, and then two critique passes run separately so visual polish could not quietly cover for functional debt.
My job in that loop was not writing prompts. It was deciding which findings were real debt and which were noise, and overriding the model when its first answer optimized for looking finished over being right. The clearest case: the second critique pass caught an operator view that had hardcoded one operator's name across every shift. The screen looked complete. A single combined review, satisfied by the polish, would most likely have waved it through.
I report what exists, which is validated demand and review results, not deployment numbers. All 3 director interviews confirmed the core problem and rejected the obvious alternative of another dashboard. The product is anchored to that.
The review held up. 2 independent passes found 4 critical and 4 major issues. Every critical was fixed before sign-off. One major was logged as documented design debt rather than dropped silently, an operator workflow that takes three taps against a two-tap spec, deferred because the real fix needs a different input method.
The structural wins are easy to state: 1 ranked entry point replaced five unranked alert sources, with ownership and deadlines built in by construction. A token-enforced design system holds all 22 screens to zero raw-value violations through a CI-ready audit. And the worker-mode decision let 2 later screens ship with no change to the original logic.
The OEE targets the product is built to move, the 5-15% gain on pilot lines and the drop in time between a risk and a response, are hypotheses to test in a pilot. I am not claiming them as results, because they are not results yet.
The thresholds in the agentic layer were set by my judgment about how costly a wrong call would be for each agent, not by usage data, because there is no usage data yet. I genuinely do not know whether a real customer's risk tolerance calls for a simpler uniform threshold instead of the tuned version I shipped. I will find out when it is in front of someone whose job depends on the answer.
The most valuable feature turned out to be the one that takes a decision off the user's plate, not the one that adds another chart. In a market racing to put more AI on the screen, the edge was the discipline to put less. The interviews are what gave me the standing to make that call, because when a director tells you to her face that she does not open the dashboard she already owns, building another one stops being a serious option.