AI Cluster
A cheap local LLM cluster built from two Mac Minis and a MacBook Pro.
The starting point was a question rather than a specification. Something in the existing setup was not working, and the first job was to establish what, which meant taking measurements before touching anything. That turned up constraints nobody had written down: a tolerance that mattered more than it looked, a material chosen for cost rather than performance, and a mounting point that could not move. Those three facts shaped everything that came after. Replace this paragraph with the actual background. What prompted the work, who it was for, and what the real constraint turned out to be once you looked closely. Around a hundred and twenty words sits comfortably at this measure without running long or feeling thin on the page.
The approach settled on the simplest thing that could be tested quickly. Rather than modelling the whole assembly up front, the first version existed only to answer one question, and it answered it in an afternoon, badly enough to be useful. The second fixed the obvious failure and introduced a subtler one. By the fourth iteration the geometry had stopped changing and attention moved to tolerances and finish. Working this way costs more prints and more bench time, but it front-loads the failures into the cheap part of the process. Replace this with your own account of how the work developed: what you tried first, what it taught you, and the decision that everything after it followed from.
Most of the effort went somewhere unglamorous. The visible parts came together quickly, while the fixturing, the wiring runs and the calibration routine took three times as long and are the reason it works reliably rather than once. A few details are worth calling out: making one component sacrificial so failures stay cheap, keeping every fastener the same size so a single tool covers the assembly, and designing in a little extra travel for adjustment. Replace this with the specifics someone interested in the craft would want. The materials, the tooling, the measurements that actually mattered, and the compromise you are least happy about having made. Photographs of a part mid-assembly usually explain more than a finished shot, so it is worth keeping the camera nearby while the work is in pieces rather than only once everything is bolted together and looking tidy.
What shipped does the job it was built for and has kept doing it. There are things a second version would change. The frame is heavier than it needs to be, the control software still assumes a setup step that could be automated, and one subassembly should have been designed for service rather than for assembly. None of those stopped it working, and all of them are on the list. Replace this with the outcome: what it does now, how it has held up, what you learned that carried over into other work, and what you would do differently having built the thing once already. It also helps to say plainly whether the result was worth the effort, since that judgement is the part a reader cannot make for themselves and is usually the reason they came looking for a write-up in the first place rather than a specification sheet.
Getting the tolerances right took longer than expected. The first assembly bound at one end and ran loose at the other, which pointed at the frame rather than the moving parts. Squaring it properly meant going back to the fixture and starting from a reference face instead of an edge. That single change removed most of the error and made the remaining adjustment small enough to do by hand. Replace this with the problem that ate the most time. Readers of build write-ups tend to find the hard part more interesting than the finished object, so it is worth being specific about what went wrong and how you found it. Include the measurement that gave it away if there was one, and say how long the whole detour took, because that is the detail people quietly compare against their own experience of the same problem.
Cost shaped several decisions. Off-the-shelf parts were used wherever the specification allowed, and the two custom pieces were designed so they could be made on equipment already to hand rather than sent out. That ruled out one obvious approach and forced a slightly awkward geometry, but it kept the whole build within budget and meant a replacement could be made in an afternoon. Replace this with the trade-offs you made and why. Constraints are usually the most useful thing to write about, because they explain decisions that would otherwise look arbitrary to someone reading it cold. Give the numbers where you can. A budget, a lead time or a machine envelope makes the reasoning concrete in a way that adjectives never manage, and it lets a reader judge whether the same approach would suit their own setup.
Testing was deliberately unkind. Rather than checking it worked once under ideal conditions, it ran repeatedly with deliberately sloppy setup to find where tolerance stacked up badly. Two failure modes appeared that had not shown up in careful use, and both were cheap to design out once known. Replace this with how you validated the work. What you measured, how many times, what you expected against what actually happened, and anything that only appeared once you stopped being careful with it. Negative results belong here too. A test that found nothing is still evidence, and saying so keeps the account honest rather than making the process look tidier and more linear than it really was on the day.
A few things would be done differently next time. Some are small, like choosing a connector that can be unplugged without removing a panel. Others are structural enough to mean starting over, which is a reasonable outcome for a first version. The useful part is knowing which is which before beginning the second. Replace this with your own reflection. It reads better at the end of a section than saved for a summary at the bottom, because it keeps the writing honest as it goes rather than tacking the caveats on afterwards. If there is a version two already sketched, this is the natural place to say what it changes and why, even briefly, since it shows the thinking did not stop when the build was finished and the photographs were taken.
The starting point was a question rather than a specification. Something in the existing setup was not working, and the first job was to establish what, which meant taking measurements before touching anything. That turned up constraints nobody had written down: a tolerance that mattered more than it looked, a material chosen for cost rather than performance, and a mounting point that could not move. Those three facts shaped everything that came after. Replace this paragraph with the actual background. What prompted the work, who it was for, and what the real constraint turned out to be once you looked closely. Around a hundred and twenty words sits comfortably at this measure without running long or feeling thin on the page.
The approach settled on the simplest thing that could be tested quickly. Rather than modelling the whole assembly up front, the first version existed only to answer one question, and it answered it in an afternoon, badly enough to be useful. The second fixed the obvious failure and introduced a subtler one. By the fourth iteration the geometry had stopped changing and attention moved to tolerances and finish. Working this way costs more prints and more bench time, but it front-loads the failures into the cheap part of the process. Replace this with your own account of how the work developed: what you tried first, what it taught you, and the decision that everything after it followed from.
Most of the effort went somewhere unglamorous. The visible parts came together quickly, while the fixturing, the wiring runs and the calibration routine took three times as long and are the reason it works reliably rather than once. A few details are worth calling out: making one component sacrificial so failures stay cheap, keeping every fastener the same size so a single tool covers the assembly, and designing in a little extra travel for adjustment. Replace this with the specifics someone interested in the craft would want. The materials, the tooling, the measurements that actually mattered, and the compromise you are least happy about having made. Photographs of a part mid-assembly usually explain more than a finished shot, so it is worth keeping the camera nearby while the work is in pieces rather than only once everything is bolted together and looking tidy.
What shipped does the job it was built for and has kept doing it. There are things a second version would change. The frame is heavier than it needs to be, the control software still assumes a setup step that could be automated, and one subassembly should have been designed for service rather than for assembly. None of those stopped it working, and all of them are on the list. Replace this with the outcome: what it does now, how it has held up, what you learned that carried over into other work, and what you would do differently having built the thing once already. It also helps to say plainly whether the result was worth the effort, since that judgement is the part a reader cannot make for themselves and is usually the reason they came looking for a write-up in the first place rather than a specification sheet.
Getting the tolerances right took longer than expected. The first assembly bound at one end and ran loose at the other, which pointed at the frame rather than the moving parts. Squaring it properly meant going back to the fixture and starting from a reference face instead of an edge. That single change removed most of the error and made the remaining adjustment small enough to do by hand. Replace this with the problem that ate the most time. Readers of build write-ups tend to find the hard part more interesting than the finished object, so it is worth being specific about what went wrong and how you found it. Include the measurement that gave it away if there was one, and say how long the whole detour took, because that is the detail people quietly compare against their own experience of the same problem.
Cost shaped several decisions. Off-the-shelf parts were used wherever the specification allowed, and the two custom pieces were designed so they could be made on equipment already to hand rather than sent out. That ruled out one obvious approach and forced a slightly awkward geometry, but it kept the whole build within budget and meant a replacement could be made in an afternoon. Replace this with the trade-offs you made and why. Constraints are usually the most useful thing to write about, because they explain decisions that would otherwise look arbitrary to someone reading it cold. Give the numbers where you can. A budget, a lead time or a machine envelope makes the reasoning concrete in a way that adjectives never manage, and it lets a reader judge whether the same approach would suit their own setup.
Testing was deliberately unkind. Rather than checking it worked once under ideal conditions, it ran repeatedly with deliberately sloppy setup to find where tolerance stacked up badly. Two failure modes appeared that had not shown up in careful use, and both were cheap to design out once known. Replace this with how you validated the work. What you measured, how many times, what you expected against what actually happened, and anything that only appeared once you stopped being careful with it. Negative results belong here too. A test that found nothing is still evidence, and saying so keeps the account honest rather than making the process look tidier and more linear than it really was on the day.
A few things would be done differently next time. Some are small, like choosing a connector that can be unplugged without removing a panel. Others are structural enough to mean starting over, which is a reasonable outcome for a first version. The useful part is knowing which is which before beginning the second. Replace this with your own reflection. It reads better at the end of a section than saved for a summary at the bottom, because it keeps the writing honest as it goes rather than tacking the caveats on afterwards. If there is a version two already sketched, this is the natural place to say what it changes and why, even briefly, since it shows the thinking did not stop when the build was finished and the photographs were taken.
The starting point was a question rather than a specification. Something in the existing setup was not working, and the first job was to establish what, which meant taking measurements before touching anything. That turned up constraints nobody had written down: a tolerance that mattered more than it looked, a material chosen for cost rather than performance, and a mounting point that could not move. Those three facts shaped everything that came after. Replace this paragraph with the actual background. What prompted the work, who it was for, and what the real constraint turned out to be once you looked closely. Around a hundred and twenty words sits comfortably at this measure without running long or feeling thin on the page.
The approach settled on the simplest thing that could be tested quickly. Rather than modelling the whole assembly up front, the first version existed only to answer one question, and it answered it in an afternoon, badly enough to be useful. The second fixed the obvious failure and introduced a subtler one. By the fourth iteration the geometry had stopped changing and attention moved to tolerances and finish. Working this way costs more prints and more bench time, but it front-loads the failures into the cheap part of the process. Replace this with your own account of how the work developed: what you tried first, what it taught you, and the decision that everything after it followed from.
Most of the effort went somewhere unglamorous. The visible parts came together quickly, while the fixturing, the wiring runs and the calibration routine took three times as long and are the reason it works reliably rather than once. A few details are worth calling out: making one component sacrificial so failures stay cheap, keeping every fastener the same size so a single tool covers the assembly, and designing in a little extra travel for adjustment. Replace this with the specifics someone interested in the craft would want. The materials, the tooling, the measurements that actually mattered, and the compromise you are least happy about having made. Photographs of a part mid-assembly usually explain more than a finished shot, so it is worth keeping the camera nearby while the work is in pieces rather than only once everything is bolted together and looking tidy.
What shipped does the job it was built for and has kept doing it. There are things a second version would change. The frame is heavier than it needs to be, the control software still assumes a setup step that could be automated, and one subassembly should have been designed for service rather than for assembly. None of those stopped it working, and all of them are on the list. Replace this with the outcome: what it does now, how it has held up, what you learned that carried over into other work, and what you would do differently having built the thing once already. It also helps to say plainly whether the result was worth the effort, since that judgement is the part a reader cannot make for themselves and is usually the reason they came looking for a write-up in the first place rather than a specification sheet.
Getting the tolerances right took longer than expected. The first assembly bound at one end and ran loose at the other, which pointed at the frame rather than the moving parts. Squaring it properly meant going back to the fixture and starting from a reference face instead of an edge. That single change removed most of the error and made the remaining adjustment small enough to do by hand. Replace this with the problem that ate the most time. Readers of build write-ups tend to find the hard part more interesting than the finished object, so it is worth being specific about what went wrong and how you found it. Include the measurement that gave it away if there was one, and say how long the whole detour took, because that is the detail people quietly compare against their own experience of the same problem.
Cost shaped several decisions. Off-the-shelf parts were used wherever the specification allowed, and the two custom pieces were designed so they could be made on equipment already to hand rather than sent out. That ruled out one obvious approach and forced a slightly awkward geometry, but it kept the whole build within budget and meant a replacement could be made in an afternoon. Replace this with the trade-offs you made and why. Constraints are usually the most useful thing to write about, because they explain decisions that would otherwise look arbitrary to someone reading it cold. Give the numbers where you can. A budget, a lead time or a machine envelope makes the reasoning concrete in a way that adjectives never manage, and it lets a reader judge whether the same approach would suit their own setup.
Testing was deliberately unkind. Rather than checking it worked once under ideal conditions, it ran repeatedly with deliberately sloppy setup to find where tolerance stacked up badly. Two failure modes appeared that had not shown up in careful use, and both were cheap to design out once known. Replace this with how you validated the work. What you measured, how many times, what you expected against what actually happened, and anything that only appeared once you stopped being careful with it. Negative results belong here too. A test that found nothing is still evidence, and saying so keeps the account honest rather than making the process look tidier and more linear than it really was on the day.
A few things would be done differently next time. Some are small, like choosing a connector that can be unplugged without removing a panel. Others are structural enough to mean starting over, which is a reasonable outcome for a first version. The useful part is knowing which is which before beginning the second. Replace this with your own reflection. It reads better at the end of a section than saved for a summary at the bottom, because it keeps the writing honest as it goes rather than tacking the caveats on afterwards. If there is a version two already sketched, this is the natural place to say what it changes and why, even briefly, since it shows the thinking did not stop when the build was finished and the photographs were taken.
The starting point was a question rather than a specification. Something in the existing setup was not working, and the first job was to establish what, which meant taking measurements before touching anything. That turned up constraints nobody had written down: a tolerance that mattered more than it looked, a material chosen for cost rather than performance, and a mounting point that could not move. Those three facts shaped everything that came after. Replace this paragraph with the actual background. What prompted the work, who it was for, and what the real constraint turned out to be once you looked closely. Around a hundred and twenty words sits comfortably at this measure without running long or feeling thin on the page.
The approach settled on the simplest thing that could be tested quickly. Rather than modelling the whole assembly up front, the first version existed only to answer one question, and it answered it in an afternoon, badly enough to be useful. The second fixed the obvious failure and introduced a subtler one. By the fourth iteration the geometry had stopped changing and attention moved to tolerances and finish. Working this way costs more prints and more bench time, but it front-loads the failures into the cheap part of the process. Replace this with your own account of how the work developed: what you tried first, what it taught you, and the decision that everything after it followed from.
Most of the effort went somewhere unglamorous. The visible parts came together quickly, while the fixturing, the wiring runs and the calibration routine took three times as long and are the reason it works reliably rather than once. A few details are worth calling out: making one component sacrificial so failures stay cheap, keeping every fastener the same size so a single tool covers the assembly, and designing in a little extra travel for adjustment. Replace this with the specifics someone interested in the craft would want. The materials, the tooling, the measurements that actually mattered, and the compromise you are least happy about having made. Photographs of a part mid-assembly usually explain more than a finished shot, so it is worth keeping the camera nearby while the work is in pieces rather than only once everything is bolted together and looking tidy.
What shipped does the job it was built for and has kept doing it. There are things a second version would change. The frame is heavier than it needs to be, the control software still assumes a setup step that could be automated, and one subassembly should have been designed for service rather than for assembly. None of those stopped it working, and all of them are on the list. Replace this with the outcome: what it does now, how it has held up, what you learned that carried over into other work, and what you would do differently having built the thing once already. It also helps to say plainly whether the result was worth the effort, since that judgement is the part a reader cannot make for themselves and is usually the reason they came looking for a write-up in the first place rather than a specification sheet.
Getting the tolerances right took longer than expected. The first assembly bound at one end and ran loose at the other, which pointed at the frame rather than the moving parts. Squaring it properly meant going back to the fixture and starting from a reference face instead of an edge. That single change removed most of the error and made the remaining adjustment small enough to do by hand. Replace this with the problem that ate the most time. Readers of build write-ups tend to find the hard part more interesting than the finished object, so it is worth being specific about what went wrong and how you found it. Include the measurement that gave it away if there was one, and say how long the whole detour took, because that is the detail people quietly compare against their own experience of the same problem.
Cost shaped several decisions. Off-the-shelf parts were used wherever the specification allowed, and the two custom pieces were designed so they could be made on equipment already to hand rather than sent out. That ruled out one obvious approach and forced a slightly awkward geometry, but it kept the whole build within budget and meant a replacement could be made in an afternoon. Replace this with the trade-offs you made and why. Constraints are usually the most useful thing to write about, because they explain decisions that would otherwise look arbitrary to someone reading it cold. Give the numbers where you can. A budget, a lead time or a machine envelope makes the reasoning concrete in a way that adjectives never manage, and it lets a reader judge whether the same approach would suit their own setup.
Testing was deliberately unkind. Rather than checking it worked once under ideal conditions, it ran repeatedly with deliberately sloppy setup to find where tolerance stacked up badly. Two failure modes appeared that had not shown up in careful use, and both were cheap to design out once known. Replace this with how you validated the work. What you measured, how many times, what you expected against what actually happened, and anything that only appeared once you stopped being careful with it. Negative results belong here too. A test that found nothing is still evidence, and saying so keeps the account honest rather than making the process look tidier and more linear than it really was on the day.
A few things would be done differently next time. Some are small, like choosing a connector that can be unplugged without removing a panel. Others are structural enough to mean starting over, which is a reasonable outcome for a first version. The useful part is knowing which is which before beginning the second. Replace this with your own reflection. It reads better at the end of a section than saved for a summary at the bottom, because it keeps the writing honest as it goes rather than tacking the caveats on afterwards. If there is a version two already sketched, this is the natural place to say what it changes and why, even briefly, since it shows the thinking did not stop when the build was finished and the photographs were taken.
