Wastewater service
Route planning and dispatch meet service records and billing in the same operational chain.
Built route planning, dispatch, service record, and billing software for a regional wastewater operator.
Standalone engineering firm / Acyclic
Private AI and operational software, built where failure has a cost.
Senior engineers work forward-deployed with your team. AI agents do much of the build; we define the acceptance criteria and stand behind what ships.
How we work
System boundary / reference architecture
Deployment pattern 01
What we build
Engineering for consequential workflows.
Our work sits close to operations, money, and sensitive information. The implementation changes by context; the accountability does not.
Private AI inside your walls
Open-weight and frontier models in a client cloud account, VPC, or hardware, with evaluation suites, inference audit trails, version pinning, and data residency designed into the deployment.
↗Software that becomes the record
Operational systems, payments, reconciliation, and field workflows where the data must survive handoffs and the edge cases cannot be ignored.
↗Forward-deployed delivery
Embedded senior engineers and agents working to explicit acceptance criteria, with a fixed outcome when the scope can be defined.
↗Research / systems evidence
Large models on small hardware.
Our own research has run 500B+ parameter open-weight mixture-of-experts models on single-workstation hardware using expert streaming and speculative decoding. We have also studied coordinated multi-model serving on shared hardware.
Read the research agendaRepresentative work
Operations have specific failure modes.
These anonymized engagements show the shape of the systems we build. No client names or outcome figures are included in this reference.
Payments and reconciliation
Money movement creates an obligation to explain each state, exception, and correction.
Built payment, reconciliation, and financial operations software around the record of what happened.
Repair orders
Labor, parts, and profitability controls need to connect to the work that produced them.
Built repair order software for automotive shops, joining service work to shop economics.
Field notes
Open questions, methods, and technical work.
Research and field notes are where we make our engineering judgment legible. These pieces are drafts.
Expert streaming for workstation-scale MoE inference
What changes when model capacity exceeds local memory, and which tradeoffs belong in the deployment decision.
An evaluation suite is a release gate
How representative cases, pinned versions, and reviewable failures make model changes manageable.
Serving multiple models on shared hardware
Coordination questions for model placement, contention, and the workloads a single machine can carry.
Acceptance criteria for operational software
Define the expected behavior before agents and engineers begin changing a system of record.
Bring us the system that has to work.
Tell us what is blocking delivery, what must be true at acceptance, and when it matters.
Start a conversation