Byline: Gregg Kell | Spotlight on Startups
Every precision aerospace shop has them — the machinists who know not to rough out the big titanium cavity too early because the part will warp, who know which tool coating will cold-weld to which alloy under which spindle conditions, who arrive at a solution before anyone else in the room can even formulate the question. They are, in Nick Khormaei‘s words, “solving problems in their heads that nobody else in the room can see.”
The trouble is, that knowledge doesn’t compound. It doesn’t transfer. And increasingly, it doesn’t stay.
A senior machinist learns, adapts, and refines over twenty years on the floor. No system captures why a specific toolpath worked on a given alloy under specific conditions. According to Deloitte’s 2025 industry outlook, 25% of the aerospace and defense workforce has more than 20 years of experience and is at or beyond eligible retirement age. Meanwhile, a 2025 APQC survey of 1,000 global organizations found that only 30% consistently capture knowledge from departing employees, while 41% rarely or never attempt it at all. When a master machinist walks out the door, that judgment goes with them. Reset. Every shop, every program, every retirement wave.
Nick Khormaei saw this firsthand as a manufacturing engineer at SpaceX and Boeing — two organizations where machining variance isn’t an operational inconvenience, it’s a mission outcome. It’s why he co-founded Neuramill, a startup building what the company calls Physical AI for Precision Manufacturing. Not a better CAM tool, but an intelligence layer that sits between engineering intent and the shop floor, encodes what the best machinists know, and makes that judgment available on every job that follows.
Commercial and military order backlogs have reached historic levels — 14,000 commercial aircraft awaiting production and a $747 billion defense backlog up 25% in two years — while talent has emerged as the single biggest constraint on execution across aerospace and defense. The challenge isn’t runway or capital. It’s the machining expertise needed to convert that backlog into delivered hardware, and it is retiring faster than it can be replaced. Khormaei’s thesis is that this is an AI problem, not just an HR problem.
The Tribal Knowledge Problem That CAM Software Was Never Designed to Solve
The global CAM market is expected to reach $5.69 billion by 2030, growing at 7.5% annually, and it is dominated by execution tools — software that translates engineering decisions into machine instructions with precision and speed. What it has never done, by design, is make the decisions.
Khormaei is direct about where the gap lives:
“CAM software does exactly what it promises. It says: make all the decisions, then input those decisions. It never sits with the machinist while they are actually making those decisions. It is a powerful execution tool. But execution is not the bottleneck.”
That bottleneck exists in the pre-CAM process planning phase — the critical, largely unstructured stage between receiving a design and opening CAM software, where a programmer must determine which features the part has, which machines and tools are valid, how tolerances interact with material behavior, and in which sequence operations should run. This is where twenty years of accumulated judgment either shows up or doesn’t.
Computer-aided process planning (CAPP) has long been recognized as the bridge between product design and manufacturing, automating tasks traditionally performed manually by experienced engineers — but CAPP systems have historically been most advantageous for large companies with high-variety, complex product geometries, leaving the high-mix aerospace shop precisely where it started: reliant on the expert who learned it the hard way and is now approaching retirement age. (For a technical overview of how generative CAPP systems reason from geometry to manufacturing sequence, the Wikipedia entry on computer-aided process planning provides useful context.)
Research from the National Institute of Standards and Technology (NIST) has estimated that poor information access costs U.S. manufacturing industries tens of billions of dollars annually in redundant work and avoidable errors. For precision aerospace shops running five-axis titanium and nickel-alloy parts, the math gets acute fast: a decision that was second nature to a twenty-year veteran now requires a junior programmer to either guess, wait for consultation time that isn’t available, or run a test cut on material that costs hundreds of dollars per pound. None of those options scale when production schedules are compressing.
The aging workforce is preparing to take decades of institutional knowledge out the door, and manufacturers are finding that fragmented data sources and disconnected systems are limiting AI’s potential — driving investment in unified data architectures that connect design and production systems. Khormaei’s argument is that the right architecture for this problem isn’t a unified dashboard — it’s a system that reasons the same way the expert does: from geometry, through physical constraints, to a manufacturable decision.
What “Physical AI” Actually Does Between Design and the Shop Floor
The term Physical AI signals a category distinction from software-native AI tools. It is AI designed to reason over the physical constraints of real-world manufacturing — geometry, tooling, material behavior, machine characteristics, and shop-floor conditions — rather than simply optimizing digital workflows or generating G-code from natural language prompts.
Khormaei describes Neuramill’s four-stage workflow in terms of the reasoning it performs, not the features it contains. The system first interprets part geometry as physical features and manufacturing intent, standardizing how similar geometry is understood consistently across programs and engineers. It then maps those features to valid operations, tools, and setups based on the specific machines, materials, and shop standards in that facility — not generic parameters, but knowledge calibrated to that shop’s actual equipment. Third, it reasons under constraints, evaluating tradeoffs across tolerance, tool access, and material behavior to ensure decisions are repeatable and grounded in real production conditions. Finally, it outputs a structured, executable manufacturing plan that machinists can read, review, and approve before a single chip is cut, including step-by-step operations with tooling and parameters, a confidence score based on similar past parts in the knowledge base, and a full audit trail in which every decision is traceable back to the geometry, the constraints, and the prior jobs that informed it.
Khormaei frames the distinction this way:
“Neuramill sits in the gap between receiving a design and entering NC code. We interpret the geometry as physical features and manufacturing intent, map those to valid operations, tools, and setups, and reason under constraints like tolerance, material behavior, and tool access to produce a structured plan with a confidence score and full audit trail. The difference is not a feature set. It is the difference between a drafting tool and a senior engineer who has seen thousands of parts and knows what is actually going to happen on the floor.”
In 2026, AI in manufacturing is no longer experimental — it has become integral to daily machine control and planning, with AI-driven approaches producing more consistent surface quality, lower tool wear, and fewer production halts. Khormaei’s specific contribution is the pre-CAM reasoning layer — the phase before any CAM tool is opened — where most manufacturing variance originates and where the least structured AI investment has historically been made.
A global survey of 300 manufacturing professionals found that while 98% of manufacturers are exploring AI, only 20% are fully prepared to implement it — with critical workflows, data flows, and exception handling remaining fragmented and manual across most facilities. The pre-CAM planning gap is a textbook example of that fragmentation: high-stakes, deeply expert-dependent, and almost entirely unstructured.
Khormaei’s internal benchmark: a complex aerospace part that would otherwise require approximately twenty hours of manual programming time, analyzed in minutes, with a structured plan ready for machinist review. At scale — across a family of related parts, across programs, across the retirement of the engineers who set the original standards — the compounding effect is the product.
Earning Trust on a Shop Floor Where AI Has a Credibility Problem
If there’s one environment where skepticism of AI is entirely rational, it’s the shop floor of a precision aerospace or defense supplier. The cost of a wrong recommendation isn’t a corrected spreadsheet. It’s a scrapped $200,000 titanium bracket, a failed first article inspection, or hardware that underperforms in a mission-critical application.
Research published in Frontiers on AI explainability in aerospace manufacturing found that transparent understanding of AI decision-making is essential for effective human-AI collaboration in environments characterized by safety requirements, system complexity, and strict regulatory adherence. That isn’t an abstract finding — it’s a precise description of why black-box AI tools fail to gain adoption on shop floors even when their outputs are technically sound. The machinist who can’t see the reasoning can’t own the outcome, and in aerospace, owning the outcome is non-negotiable.
Khormaei doesn’t argue against that skepticism. He designs around it:
“There are a lot of misconceptions about what AI will do to this industry. Being a CNC programmer is so complex, so many nuances — no AI can replace that. We are not trying to. Every plan Neuramill generates gets reviewed and approved by a machinist before anything goes to the floor.”
The trust architecture is explicit in the product. Nothing is a black box. Every output shows its reasoning — the tooling selected, the parameters chosen, the confidence score tied to how many similar parts the system has seen, and a full audit trail making every decision traceable. The machinist isn’t a bystander; they are the last gate, with everything they need to evaluate, edit, or override the output before a single chip is cut.
Khormaei reaches for the Iron Man analogy deliberately: the same expert, fully equipped, doing a lot more. The framing resonates on shop floors because it rejects the displacement narrative that makes experienced machinists resistant to AI tools, and replaces it with an augmentation argument grounded in how skilled tradespeople already think about their craft — judgment and mastery, not button-pushing.
Industry analysis confirms that the new wave of manufacturing automation is less about replacing workers and more about amplifying skilled labor — letting one technician oversee more programs, interpret analytics, and manage exceptions rather than repetitive manual tasks. The Neuramill team — combining AI scientists, former manufacturing engineers, and machinists — is built specifically to operate at that intersection, staffed by people who understand both the AI architecture and what it feels like to stand in front of a five-axis setup with a complex titanium bracket and a tight delivery window.
The audit trail carries a compliance function that matters independently of trust. ITAR compliance has tightened since 2025, with stronger documentation and traceability rules now requiring full traceability from raw material to finished component with documentation verifying conformance to specifications. Military applications impose substantially stricter requirements than commercial aerospace, including documentation retention exceeding 40 years and chain-of-custody verification across every manufacturing operation. A system that generates an inspectable, traceable record of every machining decision doesn’t just build machinist trust — it directly addresses a regulatory burden that is growing heavier as the defense industrial base reshores and scales domestically.
A July 2025 Government Accountability Office audit of foreign-supplier dependencies found that DOD’s reliance on a global network of more than 200,000 suppliers represents a mounting national security challenge — with GAO recommending tighter traceability requirements that translate into heavier documentation flow-downs to lower-tier shops. For Tier 1 and Tier 2 suppliers trying to stay ahead of that compliance curve, a system that generates AS9100-ready manufacturing records as a byproduct of its core function isn’t an add-on. It’s a structural advantage.
Why Aerospace and Defense Is the Right Beachhead — and What a First Win Looks Like
There is a counterintuitive logic to targeting the most demanding customers first. In consumer software, high tolerance for imperfection is a feature of early adoption. In aerospace and defense manufacturing, consequence is what creates genuine lock-in. Solve the problem at the hardest possible standard — where a machining error means a failed mission or a scrapped part worth hundreds of thousands of dollars — and every adjacent market gets easier.
According to International Data Corporation projections, U.S. aerospace and defense spending on AI and generative AI is expected to reach $5.8 billion by 2029 — 3.5 times higher than 2025 levels. The investment is accelerating because the production imperative is accelerating. Order books remain near historic highs, while the production ramp is constrained by bottlenecks in critical components and skilled labor, limiting how quickly backlog converts into deliveries.
Khormaei is direct about where Neuramill fits in that pressure:
“High mix work benefits the most because it requires a lot of unique setups. Aerospace and defense have these complex 3+2 and five-axis parts that have to be programmed again and again. Every job is different. Every job demands the same level of judgment. That is exactly where the leverage is highest.”
The beachhead thesis rests on two compressing pressures hitting Tier 1 suppliers simultaneously. The first is ramp-up: production, agility, and quality have become strategic battlegrounds as the aerospace and defense industrial base expands its footprint, with operating models being reset to leverage AI-enabled tools from design through production. The second is the retirement wave: PwC projects retirements will create a workforce gap of 3.5 million workers, concentrated precisely in the skilled trades and production-critical roles that precision manufacturing depends on most.
Research on AI-driven collaborative assistance in production planning found that systems capturing experiential knowledge and encoding it into structured process decisions increased the achievement of planning targets in the range of 74 to 87%. That performance range suggests the compounding model is technically grounded. The question Khormaei is running in 2026 is whether it translates from controlled research conditions to live aerospace production floors — shops that don’t have the luxury of controlled variables.
For Neuramill, an early production win looks like deploying across a family of parts at a Tier 1 supplier and producing measurable outcomes: setup time reduced, rework cut, new programmers productive faster, and an AS9100-ready audit trail satisfying quality certification requirements across the program. Once the system has ingested a shop’s process knowledge for one program, similar parts in subsequent programs benefit automatically from that encoded judgment. The relationship compounds because the product does.
Khormaei describes a consistent pattern at industry events and customer conversations: shops that arrive asking about toolpath automation, then see the full pre-CAM reasoning workflow and recalibrate what they’re actually evaluating. As he puts it:
“They come for the toolpath. They stay for the entirety of pre-CAM to toolpath.”
The programming bottleneck most shops can articulate is a symptom of the judgment-capture problem they haven’t fully named yet. Solving the named problem reveals the larger one — and by the time a shop understands the full scope of what Neuramill addresses, the conversation has moved from evaluation to deployment.
A 2026 Xometry survey of aerospace and defense leaders found that 90% cite reshoring as essential to their success — and that 60% experienced significant supplier delays in the prior year, with the industry focused on qualifying more reliable domestic manufacturing partners. A shop that can demonstrate standardized, auditable, consistent process knowledge across programs becomes a more defensible supplier in that environment — not just a more productive one.
The Open Question: Can Manufacturing AI Earn Compounding Trust?
The industrial AI landscape is full of tools that promised productivity and delivered friction. Shop floors carry the scars of pilots that never reached production, dashboards nobody checks, and systems that required more expert maintenance than they returned in value. The credibility gap isn’t irrational — it’s earned.
Neuramill’s differentiator, if the architecture holds, is that value compounds with use. Every job processed, every machinist approval, every process deviation encoded makes the next job faster and more accurate. That compounding dynamic is what separates a knowledge system from a productivity tool, and it is the right architecture for a problem that is fundamentally about knowledge accumulation, not speed optimization.
Deloitte’s 2026 A&D outlook notes that while pilot programs in AI-powered manufacturing are underway, scaling these solutions remains difficult — and full-scale industrialization is unlikely in the near term for most applications. Khormaei’s counter-argument is structural: the pre-CAM reasoning layer is one of the few places where AI can demonstrate clear, measurable value in a single program deployment, because the baseline is so inefficient, the variance from expert to non-expert judgment is so large, and the cost of getting it wrong is so visible. You don’t need twenty deployments to prove the case. You need one production floor that runs the same program twice — once before, once after — and can name the difference.
Khormaei is clear about what the 2026 production push is designed to prove:
“2026 is about proving the model in production, not in pilots. Neuramill analyses a complex part in a few minutes, whereas it would otherwise take twenty hours. We want that running on more programs, generating outcomes we can name: setup time cut, rework reduced, new machinists productive faster.”
And the human measure underneath the operational one:
“We’ve seen younger programmers stressed because they want to help but they do not want to break a part. That is kind of why we started this. The measure of success is whether that stress goes down and output goes up, visibly and verifiably.”
Khormaei isn’t measuring success by deployment count or API calls. He’s measuring it by whether the people on the floor — the programmers who understand precisely what it means to not want to break a $200,000 part — describe their work differently at the end of a year. In a market where trust is the scarce resource and the credibility gap runs deep, that is the right unit of measure.
The company has backing from investors including Creative Destruction Lab and Schema VC, and a founding team that combines Oracle AI research, SpaceX and Boeing manufacturing engineering, and patented hardware development. Support and investment inquiries are open at neuramill.co. Whether that foundation is enough to earn durable trust on the shop floors that need it most is what the next phase of production deployments will determine.
FAQ: Physical AI for Precision Manufacturing
What is Physical AI for manufacturing, and how is it different from standard CAM software? Physical AI for manufacturing refers to systems that reason over geometry, materials, machines, and physical constraints to produce manufacturing decisions — not just execute them. Standard CAM software translates decisions a programmer has already made into machine instructions. Physical AI like Neuramill sits earlier in the workflow, at the pre-CAM process planning stage, interpreting part geometry, mapping it to valid operations for specific machines and materials, and producing a structured plan with a confidence score and full audit trail. The distinction is reasoning versus translation.
How does Neuramill capture and preserve tribal knowledge from experienced machinists? Neuramill encodes manufacturing decisions — tooling selections, parameter choices, and process deviations that proved effective — as structured data tied to part geometry and shop-specific conditions. When a machinist reviews and approves a Neuramill plan, that approval becomes part of the knowledge base. Similar parts processed in the future benefit from those prior decisions automatically. The judgment doesn’t walk out the door at retirement; it stays encoded in the system and compounds with every subsequent job.
Is Neuramill designed to replace CNC programmers and machinists? No. Neuramill generates plans that machinists review and approve before anything goes to the floor. Every output is transparent and editable — the tooling, the parameters, the confidence score, and the full reasoning chain are all visible and modifiable. The design philosophy is explicit augmentation: giving skilled machinists a highly informed draft and keeping them in control of the final decision. The target outcome is one experienced machinist handling more programs with greater consistency, not fewer machinists on the floor.
What industries and part types is Neuramill best suited for? Neuramill is designed for high-mix, high-precision manufacturing in aerospace, defense, robotics, and advanced manufacturing — specifically complex 3+2 and five-axis parts that require unique setups and significant programming time for each job. These are exactly the parts where expert machining judgment is most critical, hardest to replicate consistently across personnel, and most expensive to get wrong.
Does Neuramill produce documentation that satisfies AS9100 and ITAR traceability requirements? Neuramill outputs a full audit trail in which every manufacturing decision is traceable — the tooling selected, the parameters applied, the confidence score, and the machinist approval. This structured documentation is designed to satisfy AS9100D traceability requirements and aligns with the tightened ITAR documentation standards in effect since 2025, which require full traceability from raw material to finished component with documentation verifying conformance to specifications.
How long does Neuramill take to analyze a complex aerospace part compared to manual programming? Neuramill analyzes a complex part in minutes — work that would otherwise require approximately twenty hours of manual programming time. The system interprets geometry, maps to valid operations for the shop’s specific machines and materials, and produces a structured manufacturing plan ready for machinist review in a fraction of the traditional pre-CAM to toolpath cycle.
What does a successful Neuramill deployment look like in practice? A production deployment looks like Neuramill actively running across a family of parts at a Tier 1 aerospace or defense supplier — with measurable outcomes including reduced setup time, fewer scrapped parts, faster onboarding for new programmers, and an audit trail satisfying AS9100 quality certification requirements. Success, in Khormaei’s framing, is reference customers who raise the product in conversation with other shops without being prompted — operators who can describe specifically what changed on their floor.
Nick Khormaei is the Co-Founder and COO of Neuramill, a Physical AI company building the intelligence layer for high-precision manufacturing in aerospace, defense, robotics, and advanced manufacturing. A former manufacturing engineer at SpaceX and Boeing, Khormaei holds a patent for a new test rig developed for the Boeing 777X program. Learn more about how Neuramill works, meet the Neuramill team, explore who backs Neuramill, and view open roles.
Spotlight on Startups profiles founders building companies at the edge of what’s possible. If you’re working on a problem worth telling, we want to hear from you.