Productive Machines and Manukai: Taking Machining AI from Research Lab to Shop Floor

Company Profiles

Productive Machines is a UK-based manufacturing AI company that spun out of the Advanced Manufacturing Research Center (AMRC) at the University of Sheffield in 2020 — though the underlying technology is more than a decade older. CEO and founder Erdem Öztürk spent over 10 years at the AMRC developing digital twin technology for CNC machining processes in partnership with aerospace companies including Airbus, Rolls-Royce, and the broader UK aerospace supply chain. The commercialized platform brings that research to aerospace machining suppliers as a production tool.

Manukai is a Swiss startup founded by Pascal Weber and co-founder, both of whom completed PhDs focused on applying frontier AI models to engineering and science problems. Based in the Swiss startup ecosystem centered on ETH Zurich, Manukai's approach is different from Productive Machines': rather than commercializing manufacturing-domain research, they took AI models developed for text and reasoning — the same class of models behind ChatGPT — and systematically adapted them for CNC process optimization problems.

Together, the two companies illustrate the two primary paths for manufacturing AI: deep domain research commercialized, versus frontier AI applied.

---

The Problem: Manufacturing Knowledge Has No Backup

Aerospace machining is one of the most knowledge-intensive manufacturing processes in existence. The tolerances required for turbine blades, structural airframe components, and landing gear parts are measured in micrometers. The cutting parameters — speeds, feeds, tool geometry, coolant application, fixture design — that achieve those tolerances reliably are determined by a combination of physics, empirical testing, and decades of experience.

The problem: most of that knowledge lives in the heads of machinists who have been doing this work for 30 years. It is not in process documentation. It is not in the PLM system. It is not in the CAM software. It is in the practitioner who knows that this specific alloy, on this machine, with this tool wear pattern, needs these parameters adjusted — and cannot fully explain why in a way that transfers.

This is what the manufacturing industry calls tribal knowledge. And it is leaving the workforce faster than it is being documented.

The consequences are expensive: new machinists on complex aerospace parts produce high scrap rates until they accumulate experience. Process setup for a new part takes days of trial-and-error. When a machinist retires, institutional knowledge about hard-won process parameters disappears with them.

The physics-based answer — compute the optimal parameters from first principles — is theoretically sound but computationally intractable for complex real-world machining conditions. Even with perfect material models and tool models, the number of interacting variables in a real machining operation exceeds what deterministic simulation can handle in production timescales.

---

What Productive Machines Built

Öztürk's research at the AMRC started from a specific observation: real CNC machining behavior diverges from what simulation predicts, and that divergence is systematic. Machines have specific behaviors. Tools wear in specific patterns. Workpiece materials have actual (not nominal) properties. The gap between predicted and actual is where scrap and rework live.

The Productive Machines digital twin is not a CAM simulation. It is a machine-specific, process-specific model that learns from sensor data — vibration, force, acoustic emission, spindle load — collected during actual machining operations. Over time, the model builds a representation of how this specific machine, running this specific process, actually behaves. When parameters need to be set for a new part or a new material, the twin provides parameter recommendations calibrated to the real machine rather than an idealized model.

The translation from 10 years of research to commercial product required solving problems that don't exist in a university lab:

Reliability at production rates. Research software can crash and be restarted. Production tooling cannot fail during an ongoing machining operation. The productionization work — error handling, connection resilience, graceful degradation — took significant engineering investment beyond the core algorithm.

Integration with existing systems. Aerospace suppliers run a mix of CNC controllers from Fanuc, Siemens, Heidenhain, and others, plus various MES and quality systems. Productive Machines had to build connectivity to a heterogeneous installed base rather than the controlled research environment.

Results the shop floor understands. Research outputs are optimized for scientific papers. Commercial outputs need to be actionable by a machinist or process engineer without a PhD. The UX work — translating AI recommendations into operator-facing parameters and warnings — was as important as the algorithm.

After five years as a commercial business, Productive Machines' platform is deployed in aerospace suppliers across the UK, with expansion at the Hannover Messe and international trade shows in 2025 and 2026.

---

What Manukai Built

Manukai's approach is harder to describe because it is more methodological than product-specific. Weber and his co-founder's research question was: can frontier AI models — the same underlying technology as GPT and Claude — be adapted to solve engineering problems they were not trained for?

The answer, demonstrated across their PhD work and early company applications, is yes — with the right adaptation methodology. The key insight: frontier models trained on vast amounts of text have developed reasoning capabilities that transfer to quantitative domains, including machining process optimization, more effectively than anyone expected. They do not need to be retrained from scratch on machining data. They need to be adapted, prompted, and constrained to work within the domain's rules.

Manukai's product applies this research to CNC process optimization: given a part geometry, material, machine specification, and quality requirements, what are the optimal machining parameters? This is a problem that experienced machinists solve intuitively and that CAM software addresses only partially. Manukai's AI addresses the gap — the part of the decision that currently depends on tribal knowledge.

The practical difference from Productive Machines: Productive Machines learns from your specific machine's sensor data over time. Manukai applies a reasoning model at the point of process planning, before the first cut. They are complementary, not competitive — one is real-time optimization, the other is planning-time optimization.

---

Results

Productive Machines' commercial customers report:

• Reduced trial-and-error setup time for new parts, from multiple days to hours in documented cases at aerospace suppliers
• Lower scrap rates during new part introduction, with the digital twin providing parameter recommendations that stay within the machine's reliable operating envelope
• Faster onboarding of new machinists, because process knowledge captured in the twin is transferable to operators who don't yet have 20 years of intuition

Manukai's applications, at an earlier commercial stage, demonstrate:

• Process planning cycle compression for CNC programs, with initial parameter recommendations generated in minutes for geometries that previously required experienced engineers hours to plan
• Documentation of machining rationale, creating a record of why specific parameters were chosen — a step toward making tribal knowledge explicit and auditable

Both companies participated in the Productive Machines presentation at Threaded Warwick 2026, where machining AI for aerospace was presented to an audience of UK manufacturing technology leaders.

---

Lessons Learned

1. Research takes 10 years; commercialization takes 5 more. Öztürk's AMRC research was genuinely ready for commercial use by the time Productive Machines incorporated. But "research ready" and "production deployable" are different standards, and the gap between them is real engineering work — not just packaging.

2. Frontier models transfer better than expected. Weber's finding — that LLM-class reasoning models adapt to engineering optimization problems more effectively than anyone predicted — challenges the assumption that manufacturing AI requires purpose-built domain models. The adaptation work is real, but the starting point is much higher than building from scratch.

3. The tribal knowledge problem is the market. Both companies are ultimately selling a solution to the same problem: manufacturing knowledge that lives in retiring workers' heads and nowhere else. The specific technology path (digital twin sensor learning vs. frontier model adaptation) matters less than whether you solve that problem.

4. Integration is the deployment barrier. Both companies cite machine tool connectivity — the diversity of CNC controller protocols, MES integrations, and quality system connections — as the most time-consuming part of customer deployment. The AI works. Getting the AI's inputs and outputs to flow through a 15-year-old factory infrastructure is where projects slow down.

5. University spin-outs have a credibility asset. Both companies benefit from the research pedigree of their institutions. For aerospace customers, who need to trust that a new process will not cause a safety-critical failure, a company backed by 10+ years of AMRC research and validated on Airbus programs carries more credibility than a three-year-old startup with a pitch deck.

---

Implementation Advice

If you are a aerospace machining supplier evaluating AI for process optimization, the right question is not "which AI vendor" — it is "what is your tribal knowledge capture strategy?" Productive Machines and Manukai are both valid approaches to the technical problem. The prerequisite is having enough instrumented production data to make the AI useful, and a plan for capturing the process knowledge that currently exists only in experienced machinists' heads before those machinists retire.

Start with the highest-cost tribal knowledge problem: the part families where scrap rates are highest during new part introduction, or where only one or two people in the building really know how to set them up. That is where AI ROI is most visible.

---

About the Source

This case study is drawn from AI Across the Product Lifecycle Episode 21, a podcast conversation with Erdem Öztürk (CEO, Productive Machines) and Pascal Weber (CEO, Manukai). See also: [[Digital Twin in Manufacturing]], [[CNC Machining PLM]], [[AI in Aerospace Manufacturing]], [[Knowledge Management in PLM]].