Table of Contents

Showcase

Case studies of an AI agent using HiNC through its web API on machining data anyone can fetch from a public source. Each case follows the agent as it builds a project from what was published, runs it as a simulation and checks it against pass criteria written down beforehand, then turns the result into something a machining engineer can use — the stock the program really needs, where program zero lies, cycle time, peak cutting force. Everything here is simulated: no case was cut on a real machine.

A dilemma on these pages is any point where the published data, the engine or the documentation left the agent a doubt or a choice: a missing value, a contradiction between sources, a result that looked wrong, a tool that did not behave as documented. The project the agent built is its working material and is not published; the record of how the work was managed, every dilemma met and how it was resolved, is.

Cases

Ordered from the simplest set-up to the most demanding: three-axis parts first, simultaneous five-axis last.

  • Three-Axis Milling Benchmark, Parts 1–3: From Published CAM Data to an Accepted Simulation — the first three parts of the benchmark whose parts 4–10 follow, each given as a STEP model and a CAM program: a program zero the data set never states, a stated block 1 mm too short for its own programs, a depth peak that doubles at the finer resolution with nothing wrong, and a build procedure proved by four agents that saw nothing else
  • RPTU Benchmark Parts 4–10: Seven 3-Axis Jobs Set Up and Checked by an AI Agent — seven parts from one public benchmark: a program zero that moves between parts, a stated stock that fits none of the models, a tool 5 mm shorter than the obvious choice, material the program leaves at an island's bay tips, and a check that failed for the wrong reason
  • Ball-End Finishing in Five Directions: A Force Swing That Ranks Like the Measured Roughness, and a Form Error Tool Deflection Does Not Reach — convex and concave R45 faces in 42CrMo4 from a paper that publishes only charts: the agent rebuilt both parts and wrote the ten raster programs; the swing of the cutting force along a pass ranked the directions as the measured roughness did, the force came within 3–27 % of a later measurement, the modelled tool's deflection came out a thousandth of the measured form error, and a removal-rate defect in HiNC surfaced on sloped passes
  • End Mills Run to Breakage in Hardened 42CrMo4: Seven Set-Ups Ranked against Measured Tool Life — Ø10 mm end mills run until they broke, published with no program and no tool geometry: the agent wrote the contour programs, built both cutters from their maker's catalogue and ranked HiNC's per-cycle wear against the tools' lives; the radial depth ranks as the data do per cycle but not per volume removed, the long holder is invisible to the model, and a review of the first runs caught an entry move that had faked every load peak
  • Connecting-Rod Forging Die: A Forging Rebuilt from One Drawing, a Die Sized on a Photograph, and an R1.5 Ball in an R1.5 Corner — the closed-die forging die of an aluminium connecting rod from a paper that publishes the forging's drawing and a photo of the real dies but no die drawing: a first forging overturned by three reviewer agents and a second within 0.4 % of the paper's volume, the flash land, gutter and block sized on the photo, three tools programmed on an earlier case's height-field CAM that HiNC cut to +0.04 mm of the design and timed to 0.04 %, an R1.5 ball bending ten times its limit where it finishes R1.5 corners with its whole quarter arc, two feed revisions that HiNC's step data steered and that brought its largest bend in finishing from 112.4 to 39.6 µm but not under the limit, because what is left is the entry into the corner, not the feed
  • Heatsink Porthole Die: An Extrusion Die Rebuilt from a Paper's Drawings, and Where Milling Gives Way to EDM — a two-plate extrusion die for a twelve-fin aluminium heatsink, drawn but never modelled in its paper: the agent read both plates off the figures and checked them against the paper's porthole areas, wrote four set-ups with its own CAM, HiNC caught its rapid moves grazing the roughing walls, and a Ø3 long-neck cutter's seven hours, its bending and the bearing zones it cannot reach drew the line between milling and EDM
  • OpenCellular Antenna Cover: Two Injection-Mould Inserts Derived from the Part and Milled with the Agent's Own CAM — a polycarbonate cover from an open-hardware base station, published with a drawing that states the mould's requirements but no mould: the agent derived both P20 inserts (the vent holes' draft putting 582 core pins in the fixed half, the deep bosses left to EDM), wrote every toolpath with its own CAM down to a Ø1 mm cutter in the rib slots, and played both whole inserts through acceptance within 0.4 % of its planned time; HiNC caught three mistakes in the agent's own code, read the small tools' loads too high where they meet a corner head-on, and showed an engine defect that ends a run “Finished” with nothing cut
  • Jetson AGX Thor Enclosure: A Thin-Walled Aluminium Housing in Eight Set-ups, with the Agent's Own CAM — four parts of an open-hardware computer enclosure published as models and drawings with no process: the agent wrote its own 2.5-axis CAM and two set-ups per part, a T-slot cutter reached a groove under the wall top while the wall was still backed, HiNC caught a drill moving before its length offset, a spindle asked for 3.3 times its rating and a Ø1.5 grille cutter bending 0.14 mm, and the finished parts compared with the models showed four areas no pass had reached; the accepted programs came within 1 % of the estimated 6.4 hours
  • SpaceTeamSat1 CubeSat Structure: Six Al 7075 Plates, Twenty-Four Clampings, and Whether the Small Cutters Hold Up — the six Al 7075 plates of a university team's 1U CubeSat, published as models and drawings with no process: the agent counted 24 clampings from the holes' axes, wrote its own 2.5-axis CAM and sized every holder's stick-out from a height map of the part, which a cutter held deliberately short confirmed at the place it named; the original programs overloaded the spindle and bent the D6 and D2 194 and 360 µm, and a feed scaled to the stock the cutter meets, walls finished in two passes and long cutters planned on their own brought the finishing within 25 µm on walls and 50 µm in corners but for one step; the batch takes 6.3 hours, and HiNC showed edge set-ups placed 120° about a diagonal that read no force
  • THWS Keychain: A Siemens 840D sl Program with No Model, No Stock and No Fixture — an engraving and drilling program in Siemens syntax on a five-axis machine that only positions its rotary axes: the plate read out of a drilling cycle, an engraving depth that had to come from the set-up, and 45 % of the simulated cycle found in feed moves that cut nothing
  • Mars Rover Wheel: A 0.75 mm Skin between Chevron Grousers, Milled on Four Axes — Curiosity's wheel rebuilt from published numbers and a photograph, one tread pitch milled on an A table: the skin survives only with the tool kept on a radius and a grid finer than the skin, the roughing's power peaks sat where the zig-zag pocket became a full slot and a revision slowing only those passed, and the cutter turned out to pull the skin outwards rather than press it
  • A 17-Inch Forged Aluminium Wheel: Designed from Three Public Numbers, Programmed by the Agent, Checked by HiNC Before Any Cut — a one-piece forged 6061 wheel the agent designed itself from three fitment numbers, with both set-ups written as 3+2 and five-axis toolpaths: one draft chosen so a cylindrical cutter finishes every window wall exactly, a valve hole drilled from the tire side with the head at 72.5°, HiNC's per-step power pointing at the full-width cuts of a roughing that looked clean, and HiNC's feed optimization cutting the wheel's machining time by 45 % within the spindle's rating, mostly at the feed ceilings the agent had set
  • NIST Five-Axis Test Artifacts: A Cone Frustum and a Pyramid from a Paper's Drawing — two simultaneous five-axis parts from a paper that gives only drawings and cutting conditions: the agent wrote the toolpath itself, a machine cradle too narrow for the head, a home position inside the part, and a written five-axis program whose cutting-depth peaks agree with the toolpath's to 0.00001 mm
  • BCAM Blade: Two Five-Axis Flank Finishes Told Apart Before the Cut — a thin steel blade finished with the side of a conical cutter by a commercial CAM program and by its authors' own algorithm, from six published Heidenhain programs: the agent placed the blade with the semi-finishing program alone, worked out what “a radius of 3 mm” measures, read a tool 0 and a datum shift as the NX program's cutter on a second preset, and HiNC's whole chain read the paper's 45 measured points to 0.002 mm (NX) and 0.005 mm (G1) on average — the +0.14 mm the NX program leaves already in its tool path
  • NASA HECC Centrifugal Impeller: Simultaneous Five-Axis Machining from Published Blade Coordinates — a compressor impeller published only as blade sections with no unit: the agent wrote the whole five-axis tool path itself, HiNC's cutting-load model predicted that the first path would break the tool, the owner spotted on the canvas lifts that cost a tenth of the simulated time, and HiNC's messages and a review caught retracts that grazed the part and entries that fed too fast

NC optimization cases

The same product cases taken one step further: the agent runs HiNC's NC optimization on the programs a case accepted — on some pages also on an earlier or a hand-revised version of them — as Workflow: NC Optimization describes, and replays the optimized programs on the same machine. Each page sets cycle time, spindle load, cutting force, tool deflection and the machined shape before the optimization against the same after it, measured on HiNC 3.2.45 and judged on criteria written down before the first optimization play, and says what the agent protected from the optimizer, such as finishing passes kept at their programmed feed or a floor that lets no step run slower than programmed. The pages follow the outline below and the order of the cases above.

  • Connecting-Rod Forging Die, Optimized: HiNC's Feed Optimization Takes 24.5 % off the Hand-Revised Programs, and Even the Lowest Feed Leaves the R1.5 Corner over Its Limit — the die's accepted programs and their two hand feed revisions through HiNC's feed optimization in three variants: the roughing 36.6 % shorter within the spindle-power and tool-stress targets in every variant, and the hand-revised programs, allowed only to speed up, 24.5 % shorter with their peaks unchanged; deflection targets that write 1,422 of the 1,423 over-limit steps at the corner at the 100 mm/min minimum feed bring the R1.5 ball's largest bend in steep finishing from 112.4 to 23.8 µm, still over its 12.5 µm limit, while slowing the 2 mm before and after each of those steps nearly doubles the machining time, so what is left to change is the entry into the corner or the size of the ball, not the feed
  • Heatsink Porthole Die: HiNC's Feed Optimization Takes 13 % off Both Plates, and No Feed Keeps the Ø3 Fin-Relief Cutter within 30 µm — the die's eight accepted programs through HiNC's feed optimization, each replayed on the same machine: both plates from 608.6 to 529.1 simulated minutes (−13.1 %), the porthole roughing about a third shorter at its feed-per-tooth ceiling with the spindle under a quarter of its short-term rating, and the finishing kept at its programmed chip while force targets trim the wall finishing's largest deflection from 122 to 78 µm; on the Ø3 long-neck cutter in the fin slots even the 30 mm/min minimum feed leaves 72–82 µm against the 30 µm allowed, at about 47 hours for the fin relief, the numbers behind sending it to EDM
  • Antenna-Cover Injection Mould, Optimized: HiNC's Feed Optimization Takes the Moving Insert from 12.0 to 7.9 Hours, and No Feed Holds the Ø1 mm Tool within 10 µm — the moving insert's two micro-tool programs, a Ø2 mm rest program slowed by the agent's hand corner rules and a Ø1 mm rib-slot program at the catalogue's halved slotting feed, through HiNC's feed optimization and replayed on the whole insert: 719.97 to 474.00 simulated minutes (−34.2 %), 148 of the 246 minutes saved being feed through air, the ramps and plunges the CAM had sent at the plunge feed through levels already cut, with the tools' largest stress ratio falling and the finished stock the same region by region; on a trimmed patch the optimizer found the corners without the hand rules, 20.72 against 20.61 min with them, and a target force set for the Ø1 mm tool's 10 µm limit made its trimmed program 3.8 times slower and still left a 99th percentile of 11.9 µm, so that limit is out of reach of the feed
  • Jetson AGX Thor Enclosure, Optimized: HiNC's Feed Optimization Takes 28 to 52 % off the Long-Cutter Set-ups, Holds the Cover's Finishing Bend within 25 µm, and Takes 37 % off a Grille That Looked at Its Limit — the bottom shell's and the top cover's first set-ups, two Ø12 mm cutters on the spindle's real power curve, and the back panel's Ø1.5 mm grille through HiNC's feed optimization in three variants: 197.9 to 121.7 simulated minutes for the three programs (−38.5 %), the roughing set by the chip ceiling the agent allowed rather than by the spindle; finishing held at its programmed chip in one variant, and in the other steered by a target force and a ceiling per line drawn from its own bend, which brings the cover's largest finishing bend from 36.5 to 24.2 µm and leaves the shell one wall step and 12 floor steps just over the 25 µm limit, at most 26.7 µm; and the grille cutter, whose stress looked at its target on the case's 0.25 mm grid, reaching its programmed chip on 99 % of its steps at 0.125 mm
  • SpaceTeamSat1 CubeSat Structure, Optimized: HiNC's Feed Optimization Brings the Original Program to the Spindle Target and the Deflection Limits for 3.8 % More Time, and Takes 9.9 % off the Hand-Tuned One — plate X-'s first clamping through HiNC's feed optimization at the case's 0.125 mm grid, both the program as the agent's CAM first wrote it and the hand-tuned one the case accepted: the original program's D16 roughing brought from 1.6 times the spindle's short-term rating to the 0.667 target and its D6, D4 and D2 finishing within the case's deflection limits (D2 359.9 to 49.9 µm) for 3.8 % more time, 19.9 % less than the hand-tuned program, though at chips down to 0.001 mm, so the limits hold in HiNC's model only; and the hand-tuned program 9.9 % shorter within the same limits, two thirds of it from roughing ramps raised to the full chip
  • A 17-Inch Forged Aluminium Wheel, Optimized: HiNC's Feed Optimization Takes 45 % off Both Set-ups, with the Finishing Passes Left at Their Programmed Feed — the wheel's six accepted five-axis programs through HiNC's feed optimization in three versions, each replayed on the whole wheel: optimizing every pass takes both set-ups from 17 min 25 s to 9 min 36 s (−44.9 %) within the spindle's rating but bends the finishing tool from 94.4 to 140.2 µm; with the final flank loops and the pocket floors left at their programmed feed and no step slower than programmed, 44.5 % remains and the finishing replays step for step as programmed, the saving resting mostly on the feed-per-tooth ceilings the agent chose

How a case page reads

Every case page uses the same eight headings, in this order, so two cases can be read side by side.

  1. The case — what the agent was given: the published files and what they state, and what they leave out — the stock, where the part sits, the controller, the material, the tool beyond its diameter.
  2. What the agent built — the set-up item by item — machine, fixture, stock, target model, program zero, controller, material, tool and holder, spindle, mission — each value marked as read from the source, derived from it, or chosen by the agent.
  3. How the agent managed the work — what counts as a pass, written down before the run; coarse runs before fine ones; other agents rebuilding the case from the source files and the agent's written instructions and nothing else (a blind build); other agents set to find errors in its written claims (adversarial review); sharing one server with other agents; and where a person stepped in — what the owner noticed or ruled, and what the agent changed as a result.
  4. The dilemmas — the body of the page, the small ones included. Each is told as its situation, the risk had it been missed, how it was noticed, the resolution, and the evidence that the resolution held.
  5. Results and benefits — the measured numbers with the HiNC version they were measured on; what they tell a machining engineer, a teacher or student, and someone weighing whether the approach is worth using, including what the work cost in run time and memory.
  6. Honest limits — what is assumed rather than measured, and what the engine does not check.
  7. What a reader can take to their own case — the checks that carry over to other data.
  8. Source and licence — the original source with its link and the search terms that find it again if the link moves; its licence or terms of use, named and linked; the attribution line as the licensor asks for it, including that the material is provided without warranty; what was changed from the original and what was left as delivered. This site offers no download: the reader fetches the originals from the source, and where the licence allows, the company site keeps a backup copy of the original files.

Every value on a case page says where it came from — read from the source, derived from it, chosen by the agent, or measured in a run — and measured numbers name the HiNC version they were measured on. Every picture shows the cutter in its holder at a stick-out a shop would use, because a cutter drawn hanging out of the spindle with no holder would break in practice. No case comes from a client.

The methods behind the cases

The agents build and check a case with the same public pages any caller of the web API reads: