Engineering Challenges In Manufacturing

Working with multiple LLM providers, prompt engineering, and complex data flows requires thoughtful organization. A proper structure helps teams: - Maintain clean separation between configuration and code - Implement consistent error handling and rate limiting - Enable rapid experimentation while preserving reproducibility - Facilitate collaboration across ML engineers and developers The modular approach shown here separates model clients, prompt engineering, utils, and handlers while maintaining a coherent flow. This organization has saved many people countless hours in debugging and onboarding. Key Components That Drive Success Beyond folders, the real innovation lies in how components interact: - Centralized configuration through YAML - Dedicated prompt engineering module with templating and few-shot capabilities - Properly sandboxed model clients with standardized interfaces - Comprehensive caching, logging, and rate limiting Whether you're building RAG applications, fine-tuning foundation models, or creating agent-based systems, this structure provides a solid foundation to build upon. What project structure approaches have you found effective for your generative AI projects? I'd love to hear your experiences.

5-WHY ROOT CAUSE ANALYSIS (RCA) Problem Statement: A batch of parts was rejected due to an oversized hole diameter. 5-Why Analysis: 1.Why was the batch rejected?→ Because the hole diameter was larger than the specified tolerance. 2.Why was the hole diameter too large?→ Because the drilling machine was not properly adjusted. 3.Why was the machine not properly adjusted?→ Because the operator used an outdated setup sheet. 4.Why did the operator use an outdated setup sheet?→ Because the latest revision was not available at the machine. 5.Why was the latest revision not available at the machine?→ Because there is no system in place to ensure controlled document distribution. Root Cause: No document control system for distributing updated setup sheets. Corrective Actions: •Introduce a document control procedure to issue and display the latest revision only. •Restrict access to outdated setup sheets by removing old versions from machines. •Train machine operators and line leaders on verifying document revision before setup. Preventive Measures: •Digitize all setup sheets with access through a centralized network folder or MES (Manufacturing Execution System). •Implement revision control logs with sign-off for updates and acknowledgments by operators. •Conduct regular audits on setup documents at workstations. •Establish standard work that includes a revision check step before every job setup. •Integrate barcode or QR code scanning to verify correct document versions at machines.

Don’t Automate Complexity... Simplify and Error-Proof Instead When problems arise, it’s tempting to think automation is the magic fix. But automating a broken or complex process just means you’re speeding up the production of errors. The smarter approach? Simplify the process and error-proof it (Poka Yoke) before thinking about automation. Here’s why simplification often beats automation and how you can apply it. Why You Should Simplify Before Automating: 1️⃣ Faster, Cheaper Improvements Simplifying a process through standardization and removing unnecessary steps often solves problems more quickly and at a lower cost than automation. 2️⃣ Avoid Automating Waste If your process is full of waste (like waiting, overprocessing, or rework), automating it only speeds up inefficiency. Fix the process first, then think about automation. 3️⃣ Built-In Error Proofing With Poka Yoke solutions (like jigs, fixtures, or guides), you can design processes to prevent errors from happening in the first place—without needing expensive sensors or software. 4️⃣ Flexibility and Adaptability Simplified processes are easier to adjust and improve, while automated systems can be rigid and costly to change once implemented. How to Simplify and Error-Proof a Process: ���� Map the Current Workflow: Identify unnecessary steps, bottlenecks, and areas prone to errors. ✂️ Eliminate Waste: Remove any steps that don’t add value to the product or service. 📋 Standardize Work: Create clear, repeatable instructions that everyone can follow. 🔧 Introduce Poka Yoke: Physical Error-Proofing: Use jigs, fixtures, or alignment guides to prevent incorrect assembly. Visual Cues: Use color-coded labels or visual templates to guide operators. Sensors or Alarms: Only when needed, use low-cost technology to detect errors in real time. Example of Simplification and Poka Yoke in Action: A warehouse team was dealing with frequent errors when picking products for orders. Instead of implementing a costly automated picking system, they: 1. Introduced a color-coded bin system (Poka Yoke) to help operators select the correct items. 2. Simplified the picking route to reduce unnecessary walking and waiting time. Result: Picking errors dropped by 80%, and productivity increased by 15%—all without expensive automation. When to Consider Automation: Once the process is simplified and stabilized with minimal variation, automation can enhance speed and efficiency. But it should support an optimized process, not mask its problems.

Despite the promise of 300 free electricity units a month & fat subsidies, why aren’t more Indians putting solar panels on their rooftops? Because the reality is very different from the promise. Under the PM Surya Ghar Yojana, we’re supposed to hit 40 GW of rooftop solar by 2027. But, we’re barely at 11 GW. Turns out, people still face a lot of hurdles. Subsidies take forever. Power discoms quietly resist net metering because it eats into their revenue. And after all that, a basic 3 KW system still costs around ₹70,000 even after subsidies. That’s a big ask in a country where most homeowners stretch budgets just to build a house. There’s another problem. Banks and NBFCs do offer loans for rooftop solar. But try finding someone who knows these schemes exist. Forget marketing, most people don’t even know where to begin. And yet, we desperately need rooftops to work. Because nearly 20 percent of power sent across long distances just disappears. Solar on your terrace means that electricity is generated where it’s consumed. No transmission losses. No extra burden on coal plants already gasping to meet summer demand. If we crack rooftop solar, we crack a chunk of our energy mess. That means simpler paperwork. Faster approvals. And a big push to train the workforce that actually installs and maintains these panels. We delved into this deeper in today's Finshots Link's under my profile.

Your Models Are Just 𝗘𝘅𝗽𝗲𝗻𝘀𝗶𝘃𝗲 𝗘𝘅𝗽𝗲𝗿𝗶𝗺𝗲𝗻𝘁𝘀 Without 𝗠𝗟𝗢𝗽𝘀 Most machine learning models never make it to production—or worse, they fail after deployment. Why? Because without MLOps, they remain nothing more than costly experiments. MLOps isn’t just about automation; it’s about 𝘀𝗰𝗮𝗹𝗮𝗯𝗶𝗹𝗶𝘁𝘆, 𝗿𝗲𝗹𝗶𝗮𝗯𝗶𝗹𝗶𝘁𝘆, 𝗮𝗻𝗱 𝗰𝗼𝗻𝘁𝗶𝗻𝘂𝗼𝘂𝘀 𝗶𝗺𝗽𝗿𝗼𝘃𝗲𝗺𝗲𝗻𝘁. A well-defined MLOps pipeline ensures your models don’t just work in a notebook but deliver real impact in production. Here’s the 𝗲𝗻𝗱-𝘁𝗼-𝗲𝗻𝗱 𝗠𝗟𝗢𝗽𝘀 𝗽𝗿𝗼𝗰𝗲𝘀𝘀 that transforms ML models from research to production: ⭘ 𝗗𝗮𝘁𝗮 𝗣𝗿𝗲𝗽𝗮𝗿𝗮𝘁𝗶𝗼𝗻 ✓ 𝗜𝗻𝗴𝗲𝘀𝘁 𝗗𝗮𝘁𝗮 – Collect raw data from multiple sources. ✓ 𝗩𝗮𝗹𝗶𝗱𝗮𝘁𝗲 𝗗𝗮𝘁𝗮 – Ensure data quality, consistency, and integrity. ✓ 𝗖𝗹𝗲𝗮𝗻 𝗗𝗮𝘁𝗮 – Handle missing values, remove duplicates, and standardise formats. ✓ 𝗦𝘁𝗮𝗻𝗱𝗮𝗿𝗱𝗶𝘀𝗲 𝗗𝗮𝘁𝗮 – Convert into a structured and uniform format. ✓ 𝗖𝘂𝗿𝗮𝘁𝗲 𝗗𝗮𝘁𝗮 – Organise for better feature engineering. ⭘ 𝗙𝗲𝗮𝘁𝘂𝗿𝗲 𝗘𝗻𝗴𝗶𝗻𝗲𝗲𝗿𝗶𝗻𝗴 ✓ 𝗘𝘅𝘁𝗿𝗮𝗰𝘁 𝗙𝗲𝗮𝘁𝘂𝗿𝗲𝘀 – Identify key patterns and signals. ✓ 𝗦𝗲𝗹𝗲𝗰𝘁 𝗙𝗲𝗮𝘁𝘂𝗿𝗲𝘀 – Retain only the most relevant ones. ⭘ 𝗠𝗼𝗱𝗲𝗹 𝗗𝗲𝘃𝗲𝗹𝗼𝗽𝗺𝗲𝗻𝘁 ✓ 𝗜𝗱𝗲𝗻𝘁𝗶𝗳𝘆 𝗖𝗮𝗻𝗱𝗶𝗱𝗮𝘁𝗲 𝗠𝗼𝗱𝗲𝗹𝘀 – Explore ML algorithms suited to the task. ✓ 𝗪𝗿𝗶𝘁𝗲 𝗖𝗼𝗱𝗲 – Implement and optimise training scripts. ✓ 𝗧𝗿𝗮𝗶𝗻 𝗠𝗼𝗱𝗲𝗹𝘀 – Use curated data for accurate predictions. ✓ 𝗩𝗮𝗹𝗶𝗱𝗮𝘁𝗲 & 𝗘𝘃𝗮𝗹𝘂𝗮𝘁𝗲 𝗠𝗼𝗱𝗲𝗹𝘀 – Assess performance using key metrics. ⭘ 𝗠𝗼𝗱𝗲𝗹 𝗦𝗲𝗹𝗲𝗰𝘁𝗶𝗼𝗻 & 𝗗𝗲𝗽𝗹𝗼𝘆𝗺𝗲𝗻𝘁 ✓ 𝗦𝗲𝗹𝗲𝗰𝘁 𝗕𝗲𝘀𝘁 𝗠𝗼𝗱𝗲𝗹 – Choose the highest-performing model aligned with business goals. ✓ 𝗣𝗮𝗰𝗸𝗮𝗴𝗲 𝗠𝗼𝗱𝗲𝗹 – Prepare for deployment with necessary dependencies. ✓ 𝗥𝗲𝗴𝗶𝘀𝘁𝗲𝗿 𝗠𝗼𝗱𝗲𝗹 – Track models in a central repository. ✓ 𝗖𝗼𝗻𝘁𝗮𝗶𝗻𝗲𝗿𝗶𝘀𝗲 𝗠𝗼𝗱𝗲𝗹 – Ensure portability and scalability. ✓ 𝗗𝗲𝗽𝗹𝗼𝘆 𝗠𝗼𝗱𝗲𝗹 – Release into a production environment. ✓ 𝗦𝗲𝗿𝘃𝗲 𝗠𝗼𝗱𝗲𝗹 – Expose via APIs for seamless integration. ✓ 𝗜𝗻𝗳𝗲𝗿𝗲𝗻𝗰𝗲 𝗠𝗼𝗱𝗲𝗹 – Enable real-time predictions for decision-making. ⭘ 𝗖𝗼𝗻𝘁𝗶𝗻𝘂𝗼𝘂𝘀 𝗠𝗼𝗻𝗶𝘁𝗼𝗿𝗶𝗻𝗴 & 𝗜𝗺𝗽𝗿𝗼𝘃𝗲𝗺𝗲𝗻𝘁 ✓ 𝗠𝗼𝗻𝗶𝘁𝗼𝗿 𝗠𝗼𝗱𝗲𝗹 – Track drift, latency, and performance. ✓ 𝗥𝗲𝘁𝗿𝗮𝗶𝗻 𝗼𝗿 𝗥𝗲𝘁𝗶𝗿𝗲 𝗠𝗼𝗱𝗲𝗹 – Update models or phase them out based on real-world performance. 𝘉𝘶𝘪𝘭𝘥𝘪𝘯𝘨 𝘢 𝘮𝘰𝘥𝘦𝘭 𝘪𝘴 𝘦𝘢𝘴𝘺. 𝘔𝘢𝘬𝘪𝘯𝘨 𝘪𝘵 𝘸𝘰𝘳𝘬 𝘳𝘦𝘭𝘪𝘢𝘣𝘭𝘺 𝘪𝘯 𝘱𝘳𝘰𝘥𝘶𝘤𝘵𝘪𝘰𝘯 𝘪𝘴 𝘵𝘩𝘦 𝘳𝘦𝘢𝘭 𝘤𝘩𝘢𝘭𝘭𝘦𝘯𝘨𝘦. 𝗠𝗟𝗢𝗽𝘀 𝗶𝘀 𝘁𝗵𝗲 𝗗𝗶𝗳𝗳𝗲𝗿𝗲𝗻𝗰𝗲 𝗕𝗲𝘁𝘄𝗲𝗲𝗻 𝗮𝗻 𝗘𝘅𝗽𝗲𝗿𝗶𝗺𝗲𝗻𝘁 𝗮𝗻𝗱 𝗮𝗻 𝗜𝗺𝗽𝗮𝗰𝘁𝗳𝘂𝗹 𝗠𝗟 𝗦𝘆𝘀𝘁𝗲𝗺.

What do you know about Physical AI—besides robots? (Part 2) Physical AI isn’t just about giving AI a body—it’s far more challenging than we might expect. Unlike software AI, which can be broadly applied across industries with relatively minor adaptations, Physical AI is deeply industry-specific. Since it’s tightly linked to the physical world, every industry has unique constraints and requirements that shape how AI is built and deployed. Let’s take a look at how different it can be: (the examples I chose here are all have commercial products on the market) 1️⃣ Different Industries = Different Physical Constraints Each field has its own technical, safety, and material challenges, meaning no single AI model can handle them all: - AI in prosthetics must adapt to individual users’ biomechanics and interpret neural signals in real-time. - AI in self-driving cars must process road conditions, predict human behavior, and make split-second safety decisions. 2️⃣ Hardware & Materials Are Not Universal AI-powered prosthetics & exoskeletons → Need low-latency AI chips that process movement instantly. AI-powered glasses → Require miniaturized, ultra-low-power AI that fits in lightweight frames. AI-powered infrastructure (smart traffic, self-healing roads) → Needs durable AI-embedded materials and edge computing. 💡In software AI (like LLMs), scaling mostly requires more data and better algorithms, but Physical AI needs custom-built hardware for each industry. 3️⃣ Safety & Regulation Depend on the Industry AI prosthetics & medical devices → Require FDA or CE approval to ensure patient safety. AI-powered industrial robots → Must comply with OSHA and workplace safety standards. 💡Each industry has its own regulations and risks to manage - no single "AI safety rulebook" is available. 4️⃣ Learning & Adaptation is Industry-Specific Physical AI can’t rely on generic pre-training (unlike software AI such as LLMs)—it needs industry-specific real-world adaptation. Such as: AI-powered prosthetics → Need personalized learning to adapt to each user’s unique gait. AI-powered urban infrastructure → Requires AI that adjusts to local traffic patterns, weather, and real-time events. ... This is also why Physical AI progress happens in specific industries first, rather than scaling across all industries at once like software AI. 📍If you are a leader and looking for more in-depth Executive-level AI insights, check here (10 weeks program, expertise required) ➡️https://bit.ly/3C3GSgF So, if you were to develop a Physical AI product today, which industry would you target first? Image: CB Insights _______________ For more on AI, please check my previous posts. I share my learning journey here. Join me and let's grow together. Alex Wang #artificialintelligence #machinelearning #innovation #technology

The Pentagon Just Handed American Drone Startups a $1 Billion Golden Ticket On July 10, SECDEF dropped a memo that changes everything for drone manufacturers. Combined with Trump's June 6 executive order, we're witnessing the most radical shift in defense procurement since World War II. Here's what just happened:  The Pentagon ripped up years of red tape that kept innovative companies out of defense contracts. Now they're treating small drones (under 55 pounds) like ammunition - expendable, mass-produced, and urgently needed. The numbers are staggering: • Every Army squad gets attack drones by FY2026 • Production target: Millions of units annually • Weaponization approvals: Cut from years to 30 days • Battery certifications: Down to one week For companies eyeing this opportunity, here's your roadmap: Step 1: Compliance First (Immediate) Ensure NDAA compliance - zero Chinese components. Review the Blue UAS Framework. This isn't negotiable. One foreign chip kills your entire opportunity. Step 2: Prototype Fast (12-18 months) Build modular systems under 55 pounds. Think swappable payloads for ISR or strike missions. The 18 prototypes showcased on July 17 averaged 18 months of development vs. the traditional 6 years. Step 3: Get Certified (Ongoing) Apply to DIU's Blue UAS program. This is your fastest path to approved vendor status. The memo expands this list with AI-managed updates coming in 2026. Step 4: Find Your Entry Point (30-90 days) • Respond to the Army's July 8 solicitation for low-cost systems • Partner with established primes as a subcontractor • Target frontline units are now empowered to buy directly Step 5: Scale Smart (By 2026) Secure private funding. Explore DoD purchase commitments. Participate in the new drone test zones launching in 90 days. The brutal reality? We're playing catch-up. China produces 90% of commercial drones globally. But that's precisely why this opportunity exists. The Pentagon needs American manufacturers desperately. Watch for these challenges: • Supply chain constraints for non-Chinese components • Fierce competition from AeroVironment and Kratos • Higher production costs vs. Chinese competitors • Maintaining cybersecurity while moving fast Stock prices tell the story - drone companies surged 15-40% after the announcement. Private capital is flooding in. America is building a new arsenal, and drones are the foundation. If you have manufacturing capability, AI expertise, or can build at scale, this is your Manhattan Project moment. The difference? This time, we know exactly what we're building and why. The window is open. But it won't stay that way.

A month ago, I shared a simulation video of the 𝐃𝐢𝐫𝐞𝐜𝐭𝐞𝐝 𝐄𝐧𝐞𝐫𝐠𝐲 𝐃𝐞𝐩𝐨𝐬𝐢𝐭𝐢𝐨𝐧 (𝐃𝐄𝐃) process of a titanium wire. Since then, we've added 𝐆𝐏𝐔 𝐬𝐮𝐩𝐩𝐨𝐫𝐭 to our simulation software, significantly reducing simulation time and enabling more complex and detailed studies. The following video demonstrates the deposition process of a titanium wire (1 mm radius) on a (4 x 4) cm² substrate across 𝐟𝐨𝐮𝐫 𝐥𝐚𝐲𝐞𝐫𝐬. The wire is melted using 𝐭𝐡𝐫𝐞𝐞 𝐆𝐚𝐮𝐬𝐬𝐢𝐚𝐧 𝐥𝐚𝐬𝐞𝐫 𝐛𝐞𝐚𝐦𝐬, each's power is individually controlled to maximize the deposition rate. The breakage of the liquid bridge connecting the wire and substrate can be observed during the deposition of the last track in the fourth layer. We employ a 𝐫𝐚𝐲 𝐭𝐫𝐚𝐜𝐢𝐧𝐠 algorithm to model the laser-material interaction, where the total laser power is distributed among numerous rays. The laser energy absorbed by the material surface is computed based on ray intersections with the material surface, considering surface temperature, angle of incidence, and polarization. In the video, the upper section displays the temperature field, while the lower section shows the number of ray intersections with the material surface throughout the simulation. Simulated on a Ryzen 7950x3D and an RTX4070. The video is rendered using Blender. Get in touch with us at blank-simulations if you see potential application scenarios. #SPH #multiphysics #raytracing #additivemanufacturing 𝐌𝐞𝐭𝐡𝐨𝐝: - Smoothed Particle Hydrodynamics (SPH) - MPI-OpenMP parallelization - GPU-acceleration - Dynamic workload balancing - Adaptive particle refinement 𝐏𝐡𝐲𝐬𝐢𝐜𝐬: - Ray tracing to model laser-material interaction - Temperature-dependent material properties - Latent heat of fusion and crystallization - Evaporation and recoil pressure - Surface tension and wetting

Conifer's huge $20M seed round just unlocked a smart way around our rare earth dependence. Every EV needs motors, but we rarely talk about them. Most discussions (and funding) focus on batteries, while motors remain tied to China's rare earth monopoly. Conifer's team flipped this by developing electric hub motors using ferrite magnets instead of rare earths. Simple switch, massive implications: The motors deliver 10% better range while being half the size of competitors. What is their smart move? Building automated production lines near customers - no massive factories, just local microfactories cranking out motors. For manufacturers, it's literally plug-and-play. It's exactly the kind of climate tech we need more of: Better performance, simpler supply chains, easy adoption. Sometimes the biggest impact comes from rethinking the basics rather than chasing the next breakthrough. Any hardware founders working on overlooked EV components? Drop a comment.

Manufacturing processes are often plagued by inefficiency.   Here's why:   Manufacturers cling to old batch habits. ___   Batch Production is a traditional manufacturing method where identical or similar items are produced in batches before moving on to the next step.   Some manufacturers argue that large batches balance workloads and minimize changeovers.   But data often shows otherwise.   Overlong production runs cause overproduction. Operators lose focus working on large batches while equipment drifts out of standards between changeovers.   Main drawbacks:   -Piles of WIP inventory waiting for the next step -Defects hide among the batches -Inefficient space management -Uneven workflow -Long lead times   Those lead to:   -Some stations being overloaded, others waiting -Low responsiveness to customer demand -More scrap and rework -Higher carrying costs -Facility costs up   Switching to One-Piece Flow can bring relief.    Workstations are arranged so that products can flow one at a time through each process step, making changeovers quick and routine.   Main advantages:   +High customer responsiveness +Minimal work-in-process inventory +Quality issues are detected immediately +Reduced wasted space and material handling +Easy to level load production to match takt time   The selection between batch processing and one-piece flow can significantly impact quality, productivity, and lead time in a manufacturing process.   P.S. Some case studies show improvements in labour productivity of 50% or more. Lead times can drop by 80%. And quality can approach Six Sigma.