TightLoop neuromorphic control engine
TightLoop learns the shake, drift, tremor, overshoot, and contact error left by existing controllers, then turns those residual patterns into adaptive neuromorphic correction for the next control output: steadier view, steadier hand motion, longer lock-on, softer contact, and cleaner precision behavior.
What TightLoop is
TightLoop adds neuromorphic residual-error control to an existing controller. It learns what still remains after normal control and helps turn it into steadier view, longer target hold, softer contact, fewer visible corrections, and clearer validation data. Available as SDK, runtime module, or OEM license.
Why it is different
TightLoop is different because it combines three ways of seeing residual error in one control loop: fast adaptation, history-shaped state geometry, and selective correction of recurring failure sectors. It is built for residual patterns that simple averaging, fixed gains, or ordinary smoothing can leave behind.
The adaptive core. Spike timing, heterogeneous time constants, and reward-modulated plasticity let the controller react while the signal is still changing. The advantage is fast adjustment to wobble, tremor, vibration, and contact shifts without waiting for a slow retune.
The memory geometry. Past inputs are compressed into a state-function landscape instead of a longer buffer. Inspired by reservoir computing, power-law memory, neodymium-like hysteresis, and transcendental-number-like non-repetition, it helps separate similar residual regimes from truly different ones.
The final error-correction layer. Ideas from quantum error correction and holographic-wedge geometry are used as inspiration for narrowing a wide error space into logical or failure sectors. HRR does not cover the base controller; it targets the recurring misses that remain near the final command.
Product lines
TightLoop applies wherever residual error reduces product value: Game/VR adaptive cameras, robotic surgery, teleoperation, EO/IR turrets, precision optics, grippers, gimbals, drones, and sensor heads.
A TightLoop SDK path for developers who need faster tracking, steadier horizon, operator-view stabilization, and replayable before/after motion. Unity receives a native runtime, C# camera component, sample scene, and motion-record exporter. Unreal/Fab receives a plugin path with Blueprint/C++ components and sample camera rigs.
vr.cording.aiFor remote surgery and robotic surgery workflows, TightLoop targets screen motion, tool tremor, overcorrection, and command jitter. The first deployments fit training systems, simulators, research teleoperation, and device-evaluation software before regulated clinical integration.
surgery.cording.aiFor EO/IR cameras, pan-tilt heads, and long-range inspection systems, TightLoop reduces the line-of-sight motion that remains after strong baseline control. The customer value is clearer sightline, less visible drift, and longer target hold.
For inspection, metrology, precision stage, and optical stabilization equipment where residual error affects yield, alignment, or measurement quality.
For robot grippers and contact-rich tools where slip must fall without raising overgrip and damage risk.
For drones, payload gimbals, mobile inspection rigs, OIS, and robot sensor heads that need steadier image or pointing behavior.
Buy and deploy
Send a short sample of controller input, controller output, and measured product result. cording.ai will propose the scope, validation method, price, and the right path to SDK, runtime module, or OEM license.
Target value, reference position, user command, tracked object position, current feedback, or current state value.
Motor command, speed command, position command, torque command, correction command, or camera, gimbal, tool, and turret drive command.
Actual position, actual angle, velocity, image or view displacement, distance from target, sightline offset, contact force, pressure, or slip amount.
Files can be in any convenient readable format. One time-aligned table is best, but multiple timestamped files are also fine. CSV, TSV, JSON, Parquet, MAT, XLSX, TXT, or readable binary exports are acceptable. Example columns: time_s, controller_input_*, controller_output_*, measured_result_*, target_optional, error_optional
Send the metric you want to improve, plus any comments or questions, together with the data files by email.
A paid entry product that reviews supplied control data, target error, validation scope, and the next commercial step.
A paid validation product for HIL, simulator, bench, or field testing with before/after performance criteria.
A deployable TightLoop SDK, runtime module, or OEM control module for integration into customer products.
Current measured signals
The numbers below are internal replay and fixture results. Customer deployment still requires customer logs, simulator evaluation, bench testing, and domain-specific validation. Check whether the metric that needs improvement actually improves.
Average tool-motion error reduced from 0.001010 to 0.000407; large spikes reduced 5.05x.
Operator-view replay showed a 3.53x lower horizon error signal and 1.88x lower operator-view error.
Standard and stress turret tests lowered remaining line-of-sight error against an upgraded robust controller.
Slip RMS improved 1.43x, while physical overgrip averaged 3.01x lower and audited damage mean was 2.63x lower.
Industrial IMU fixture result: RMS lowered from 0.004149 to 0.001427 with adaptive assist.
Standard and stress optical tests lowered residual error against an upgraded optical controller.
Research archive
cording.ai keeps earlier papers, datasets, and videos visible because they show the technical path that led to TightLoop: time-reactive control, structured error correction, entropy, camouflage, fusion-edge instability, and motion stabilization.
The early control idea: a spiking liquid-style neuromorphic system built around spike timing, online adaptation, and learning from residual error.
Open ZenodoA BOUT++ fusion-edge instability experiment where neuromorphic runs delayed runaway stress from simulation time 50 to 72-76 in eight of nine runs.
Open ZenodoA memory-geometry concept where past inputs shape a compact state-function landscape, giving residual errors multi-scale and attractor-like structure.
Open ZenodoA structured error-correction experiment connected to logical-sector proposals, quantum error correction, and holographic wedge geometry.
Open ZenodoA feedback-time theory for event-driven systems, attention weight, relation-driven meaning, and prediction difficulty in dense feedback loops.
Open ZenodoA software-defined neuromorphic entropy experiment that showed credible randomness with 186/188 NIST STS subtests passed.
Open ZenodoA multiband active-camouflage concept where adaptive control adjusts visible and thermal tile targets to reduce contrast and tracking quality.
Open VideoA short stabilization record connected to view orientation, OIS, drone payloads, and robot sensor heads.
Watch VideoA hand and tool-control demonstration showing why tremor correction matters for human-machine operation.
Watch VideoA control and maintenance experiment connecting PID tuning, anomaly detection, and predictive maintenance signals.
WatchTightLoop Sentinel
TightLoop Sentinel is an early-warning product for predictive maintenance. It watches plant and equipment data and turns weak warning signs into clear maintenance alerts before downtime, rejected product, overheating, or wear grows. It can strengthen the predictive maintenance tools you already use, and it can also be tested as a standalone predictive maintenance product.
Customer offer
Connect historical sensor data, alarm history, inspection notes, and maintenance records. Sentinel looks for patterns that usually appear before failure, quality loss, overheating, wear, clogging, or shutdown.
Your team sends operating logs and known incidents. cording.ai shows how much earlier Sentinel would have warned, how many false alerts it would add, and which equipment area needs attention.
You receive before and after alert timing, missed problem checks, estimated business impact, and a simple video your operations and executive teams can review.
Why customers buy
One earlier maintenance window can protect production, reduce emergency labor, avoid rejected product, and keep expensive equipment running. Sentinel can start beside your current maintenance software, then become a standalone option if the pilot proves stronger performance.
These examples are simulation results. Real deployment starts with your data and a check at each site.
In a vial filling example, Sentinel warned 12.1 minutes earlier. At high line speed, that can mean thousands fewer units exposed to a hidden filling problem.
+12.1 min Watch on YouTube Data center coolingIn a liquid-cooling example for an artificial intelligence data center, Sentinel warned 36.4 minutes earlier about cooling-loop fouling while keeping false alerts unchanged in the simulation.
+36.4 min Watch on YouTube Wind power maintenanceIn a wind turbine gearbox example, Sentinel opened a 15.95-day maintenance window, giving operators more time to plan service before failure risk grows.
+15.95 days Watch on YouTube Healthcare ICUIn an ICU monitoring simulation, Sentinel reduced unnecessary alarms from temporary signal noise and warned 113.6 minutes earlier when the patient was getting worse.
-7.538%p Watch on YouTube Turbofan maintenanceIn a NASA turbofan benchmark, a neuromorphic warning layer worked beside a deep life-prediction model to give earlier maintenance warnings with controlled false alerts.
+7.45 cycles Open Zenodo Predictive maintenanceA Zenodo paper summarizes public tests on jet engine life prediction, bearing failure, and wind gearbox vibration, showing how earlier warnings and fewer false alerts can work with existing maintenance tools.
3 benchmarks Open Zenodo Edge forecast-to-actionA Jetson Orin replay study shows how TightLoop Sentinel can turn zero-shot forecast uncertainty into practical reserve, alert, and capacity actions for operations teams.
4 datasets Open ZenodoContact
Contact.
Founder: Kyuchul Lee, based in Daegu, South Korea. MD, family medicine specialist, Master of Medical Science, MBA, software developer, meditator, and investor.