Work

The programs, and what they changed.

We do not publish client names or logos. The organizations we work with operate in defense, in regulated medicine, in industry, and in critical infrastructure, and discretion is part of the job. What we can show is the shape of each program: what was at stake, which layer the hard decision was actually made at, and what changed as a result.

Eleven engagements, ordered to show the span rather than the chronology. Four of them have no model in them anywhere. Two run on hardware that could not be taken out of service. One is enforced beneath a flight controller.

Infrastructure & Fleet PlatformsArchitect + Build

One control plane over a mixed-vendor server fleet across three out-of-band protocols, where the hard problem is detecting the absence of a signal.

25,000nodes, three protocols, one API

Defense & Uncrewed SystemsArchitect + Harden

A spoken command from a handheld radio becomes a machine action, with no transcription in the path and a default under any uncertainty of doing nothing.

13.4 msaudio to command, 22 ms budget

Medical DevicesArchitect

A high-stakes biometric wearable architected end to end, where the physics sets a recoverability ceiling months before there is a dataset to test it against.

3disciplines, one owner

Industrial AutomationBuild + Harden

Adaptive real-time control on a hydraulic top-drive system, over Modbus TCP to a live PLC through an industrial VPN gateway, on equipment that could not be taken out of service.

liveequipment, no model

Distributed SystemsDiagnose + Harden

A product that ran fine at low load and buckled as it grew, where adding servers had stopped helping and was making it worse.

74%of request time, in the database

Robotics & Autonomous SystemsBuild + Harden

Perception and control as one real-time loop closing through the physical world, built and tested with hardware in the loop, inside the control loop's deadline.

sim-to-real distribution · jitter headroom budgeted · hardware-in-the-loop

Industrial AutomationArchitect + Build

A condition-monitoring system across an 87-unit industrial fleet, tuned to the cost of being wrong rather than to accuracy, so the crew believes the alerts that remain.

41%fewer false alarms

Medical DevicesDiagnose + Harden

Pre-manufacturing firmware for a patient-safety neuro-wearable, where the defect that mattered was one layer below the assignment and days from being permanent.

pre-freezecaught before the BOM locked

Governed AI & RetrievalArchitect + Operate

A product answering questions over governed data, where trust is a property of the system around the model rather than of the model itself.

abstainis a valid answer

Document IntelligenceDiagnose + Build

An extraction system at 44% accuracy, low enough that a person re-checked nearly everything, rebuilt so the system knows which cases it is unsure about.

44 to 95%in five days

Enterprise SoftwareArchitect + Operate

Agents doing real operational work, where the field competes on whether you can build them and the real question is whether they survive production.

1.9Mtool calls, under 0.4% hallucination

AI InfrastructureArchitect + Operate

A platform at 60,000+ requests a day where the accuracy was fine and the bill was not, because near-identical work was recomputed every time.

58%lower monthly cost

Machine VisionArchitect + Harden

A camera migrated from a proprietary transport to a standard one, sensor and FPGA untouched, where the failure that decides the product is silent.

10 Gbpspreserved through DMA

Firmware & OTABuild + Operate

Over-the-air updates on a no-downgrade platform, where a clean staged rollout of a bad build still bricks its percentage of the fleet.

2%where the release was held

Computer VisionArchitect + Operate

A visual judgment automated, where the deliverable was not which pixels should change but proving which pixels must never change, on every image.

zerounexpected changes, per image

Biosignal MLModel + Operate

Biosignal systems where an accuracy number inflated by subject-level leakage is worse than no number, because it drives a claim that does not hold.

3modalities, subject-disjoint

AI InfrastructureArchitect + Build

Private per-customer similarity over an API, where the comparison engine is the easy part and isolation and run cost decide it.

zerocross-tenant leaks

Industrial IoTArchitect + Operate

Remote monitoring for hardware whose controller may be undocumented or encrypted, with two paths through the first gate.

2paths, one owner

Industrial IoTDiagnose + Operate

Field sensors to a cloud dashboard, where the risk was capital rather than capability: own one hop and you ship a dashboard that lies politely.

end to endproven before fleet commit

Fintech & RiskArchitect + Operate

An advisor-facing product where the posted need read as throughput and the real need was that every number is fresh, entitled, and never fabricated.

every figuretraced to source

Verification & ComplianceArchitect + Operate

Verifying real-world data across defended portals, where scraping had already failed and the authorized path was the only one that survives success.

authorizedpath first

MechatronicsArchitect + Build

A mechanism on a finished garment that had to look alive, where a convincing beat is asymmetric and the hard part moved into the controller.

hoursnear-silent on a body

Embedded & BLEDiagnose + Build

Firmware on a constrained BLE part, where negligible impact is not a promise about the sensor but a guarantee about the topology.

0.04microamp-hours per read

The upper layers

Four engagements at the model and serving layer, framed as what they are.

These sit at the top of the stack, and they are production reliability problems rather than model-building ones. Most of what makes them hard is old distributed-systems discipline arriving in new clothing: drift detection, regression, idempotency, and knowing when the system does not know.

Read as a set

Four of these have no model in them, and that is the point.

Any firm can publish a list of projects. What is worth reading here is the shape of the set. The fleet platform, the hydraulic control, the scaling work, and most of the predictive-maintenance architecture are systems, protocol, and signal engineering. A vendor whose only product is a model would have found a use for a model in every one of them.

The other thing worth noticing is the vertical span. The same principal enforced a fail-closed gate beneath a flight controller, scheduled polling against baseboard-controller session limits, and designed a biosignal chain against a physical recoverability ceiling. Those layers are usually owned by different people. The value of one owner is that the decisions which have to be made together get made together.

What we hold back

The technical specifics are the point. The clients are not.

Every account here is anonymized. The protocols, the silicon, the standards, and the architecture decisions are named because they are what a technical evaluator needs to judge whether the engineering is real, and none of them identify a client.

A portion of what we do sits under agreements that do not allow even this much. If you need to evaluate work closer to yours before committing a program, we can speak to it directly under the right terms.

How we engage

Your program is not on this page.
Tell us what it is.

The work closest to yours is often the work we cannot publish. Describe the system and the stakes and we will tell you what is relevant, and what we would own.

Start a conversation

mostafa@opulion.dev · Response within 24 hours · By inquiry

The programs, and what they changed.

One control plane over a mixed-vendor server fleet across three out-of-band protocols, where the hard problem is detecting the absence of a signal.

A spoken command from a handheld radio becomes a machine action, with no transcription in the path and a default under any uncertainty of doing nothing.

A high-stakes biometric wearable architected end to end, where the physics sets a recoverability ceiling months before there is a dataset to test it against.

Adaptive real-time control on a hydraulic top-drive system, over Modbus TCP to a live PLC through an industrial VPN gateway, on equipment that could not be taken out of service.

A product that ran fine at low load and buckled as it grew, where adding servers had stopped helping and was making it worse.

Perception and control as one real-time loop closing through the physical world, built and tested with hardware in the loop, inside the control loop's deadline.

A condition-monitoring system across an 87-unit industrial fleet, tuned to the cost of being wrong rather than to accuracy, so the crew believes the alerts that remain.

Pre-manufacturing firmware for a patient-safety neuro-wearable, where the defect that mattered was one layer below the assignment and days from being permanent.

A product answering questions over governed data, where trust is a property of the system around the model rather than of the model itself.

An extraction system at 44% accuracy, low enough that a person re-checked nearly everything, rebuilt so the system knows which cases it is unsure about.

Agents doing real operational work, where the field competes on whether you can build them and the real question is whether they survive production.

A platform at 60,000+ requests a day where the accuracy was fine and the bill was not, because near-identical work was recomputed every time.

A camera migrated from a proprietary transport to a standard one, sensor and FPGA untouched, where the failure that decides the product is silent.

Over-the-air updates on a no-downgrade platform, where a clean staged rollout of a bad build still bricks its percentage of the fleet.

A visual judgment automated, where the deliverable was not which pixels should change but proving which pixels must never change, on every image.

Biosignal systems where an accuracy number inflated by subject-level leakage is worse than no number, because it drives a claim that does not hold.

Private per-customer similarity over an API, where the comparison engine is the easy part and isolation and run cost decide it.

Remote monitoring for hardware whose controller may be undocumented or encrypted, with two paths through the first gate.

Field sensors to a cloud dashboard, where the risk was capital rather than capability: own one hop and you ship a dashboard that lies politely.

An advisor-facing product where the posted need read as throughput and the real need was that every number is fresh, entitled, and never fabricated.

Verifying real-world data across defended portals, where scraping had already failed and the authorized path was the only one that survives success.

A mechanism on a finished garment that had to look alive, where a convincing beat is asymmetric and the hard part moved into the controller.

Firmware on a constrained BLE part, where negligible impact is not a promise about the sensor but a guarantee about the topology.

Four engagements at the model and serving layer, framed as what they are.

Finished, or merely stopped

Knowing when it does not know

A composed script, at volume

Four of these have no model in them, and that is the point.

The technical specifics are the point. The clients are not.

Your program is not on this page.
Tell us what it is.

The programs, and what they changed.

One control plane over a mixed-vendor server fleet across three out-of-band protocols, where the hard problem is detecting the absence of a signal.

A spoken command from a handheld radio becomes a machine action, with no transcription in the path and a default under any uncertainty of doing nothing.

A high-stakes biometric wearable architected end to end, where the physics sets a recoverability ceiling months before there is a dataset to test it against.

Adaptive real-time control on a hydraulic top-drive system, over Modbus TCP to a live PLC through an industrial VPN gateway, on equipment that could not be taken out of service.

A product that ran fine at low load and buckled as it grew, where adding servers had stopped helping and was making it worse.

Perception and control as one real-time loop closing through the physical world, built and tested with hardware in the loop, inside the control loop's deadline.

A condition-monitoring system across an 87-unit industrial fleet, tuned to the cost of being wrong rather than to accuracy, so the crew believes the alerts that remain.

Pre-manufacturing firmware for a patient-safety neuro-wearable, where the defect that mattered was one layer below the assignment and days from being permanent.

A product answering questions over governed data, where trust is a property of the system around the model rather than of the model itself.

An extraction system at 44% accuracy, low enough that a person re-checked nearly everything, rebuilt so the system knows which cases it is unsure about.

Agents doing real operational work, where the field competes on whether you can build them and the real question is whether they survive production.

A platform at 60,000+ requests a day where the accuracy was fine and the bill was not, because near-identical work was recomputed every time.

A camera migrated from a proprietary transport to a standard one, sensor and FPGA untouched, where the failure that decides the product is silent.

Over-the-air updates on a no-downgrade platform, where a clean staged rollout of a bad build still bricks its percentage of the fleet.

A visual judgment automated, where the deliverable was not which pixels should change but proving which pixels must never change, on every image.

Biosignal systems where an accuracy number inflated by subject-level leakage is worse than no number, because it drives a claim that does not hold.

Private per-customer similarity over an API, where the comparison engine is the easy part and isolation and run cost decide it.

Remote monitoring for hardware whose controller may be undocumented or encrypted, with two paths through the first gate.

Field sensors to a cloud dashboard, where the risk was capital rather than capability: own one hop and you ship a dashboard that lies politely.

An advisor-facing product where the posted need read as throughput and the real need was that every number is fresh, entitled, and never fabricated.

Verifying real-world data across defended portals, where scraping had already failed and the authorized path was the only one that survives success.

A mechanism on a finished garment that had to look alive, where a convincing beat is asymmetric and the hard part moved into the controller.

Firmware on a constrained BLE part, where negligible impact is not a promise about the sensor but a guarantee about the topology.

Four engagements at the model and serving layer, framed as what they are.

Finished, or merely stopped

Knowing when it does not know

A composed script, at volume

Four of these have no model in them, and that is the point.

The technical specifics are the point. The clients are not.

Your program is not on this page.Tell us what it is.

Your program is not on this page.
Tell us what it is.