Systems engineered
to specification.
Cognitive Horizon is an engineering consultancy that draws the system before committing it to steel. We are retained when the project is too load-bearing to be built by guesswork — when a mistake at the foundation is a mistake at the roof.
Five disciplines, drawn to
the same tolerance.
AI / ML Engineering
Evaluation harnesses, agentic systems, RAG, fine-tuning, MLOps — built to degrade gracefully.
- LLMs
- Agents
- RAG
- Evals
Application Development
Web, mobile, desktop, CLI — the right language for the problem, at a bar that holds from line one.
- Rails
- React
- Phoenix
- SwiftUI
Systems & Infrastructure
Rust, Go, C/C++ services. Distributed systems engineered — not assembled.
- Rust
- Go
- Kafka
- K8s
Algorithm Design
Competition-grade rigour applied to graph, geometry, numerical and combinatorial problems.
- Graphs
- Geometry
- Numerical
Consulting & Architecture
Technical audits, architecture reviews, fractional CTO — judgement on the calls that matter most.
- Audits
- Fractional CTO
Drawn, then
built.
Every engagement follows the same order of operations — because drawing a load-bearing wall in the wrong place is much cheaper than framing it.
Survey the site.
We read your code before we write ours. What constraints bind the problem — regulatory, legacy, organisational, economic? Which walls can come down, which cannot?
Draw the system.
A dimensioned proposal: interfaces, data shapes, failure modes, cost envelopes. Two pages of clear drawing beats twenty pages of vague intent.
Build to specification.
Implementation against an approved drawing. Variances are flagged in red-pencil revision marks — never buried in a PR description.
Sign and date.
Every drawing is signed. We stand behind the systems we draw. Post-launch, we are the number you call at 3 AM when the load exceeds its rating.
Real-time analytics pipeline
Millions of events per second, sub-millisecond p99 latency. Rust service core.
- Rust
- Kafka
- Postgres
Document intelligence platform
LLM-based system; manual review reduced by ninety percent across three operating units.
- LLMs
- RAG
- Python
A problem worth
dimensioning?
Tell us the scope in a paragraph, the timeline in a line, and the names of the people we would be working with. We reply within a business day with a judgement, not a pitch.
Observation
Every system has two drawings —
the one you sketched before it
was built, and the one you wish
you had sketched after.
— we make the first one count.