Carraggon DSES — Domain-Specific Expert Systems
Frontier models are trained on the internet. DSES trains models on your company — your documents, your tools, your judgement calls and your outcomes — using reinforcement learning, until they perform your highest-value work at expert level.
Trained on
Your work, not the web
Method
RL against real outcomes
Benchmark
Your experts, your rubric
Deployment
Your cloud or ours
The limit isn't the model. It's that no model has ever done your job, in your company, and been told whether it got it right.
The approach
01
A model that does one job in your business exceptionally well.
General models know a little about everything and nothing about how your firm actually operates. DSES builds a model per high-value function — underwriting, diligence, claims, coding, support, trading ops — trained until it matches the people who do it best.
02
Reward the result the business actually wants.
We define a reward from your own graded outcomes — approved deals, clean audits, resolved tickets, accepted drafts — then run RL loops until the model reliably produces them. Expert feedback is captured in the loop, not in a document nobody reads.
03
The model practises in a copy of your world before it touches production.
We build a sandboxed environment with your tools, documents, APIs and guardrails so the model learns to use them — not just to talk about them. Every rollout is scored, replayable and safe.
04
If it doesn't beat your team's baseline, it doesn't ship.
Every DSES model is measured on a rubric your experts wrote, against a held-out set of your real work. You see accuracy, cost, latency and failure modes before deployment — and a live scoreboard after.
Domains
Private equity & finance
Diligence memos, CIM extraction, covenant review, portfolio monitoring and IC-ready analysis.
Insurance
Submission triage, underwriting judgement, claims adjudication and fraud signal detection.
Legal
Contract review against your playbook, obligation extraction, precedent search and drafting.
Healthcare
Coding and documentation, prior authorisation, care-pathway review and clinical summarisation.
Real estate
Lease abstraction and audit, application screening, maintenance triage and owner reporting.
Software & engineering
Codebase-specific agents that follow your conventions, tests, review standards and release process.
Support & service
Resolution agents trained on your best reps' transcripts, tooling and escalation rules.
Defence, space & industrial
Requirements analysis, compliance review and operational decision support in restricted environments.
How it works
01
Scope the job
We pick one function where expert judgement is the bottleneck, and write the rubric with the people who own it.
02
Build the environment
Your tools, data and guardrails are wired into a sandbox where the model can practise the full task end to end.
03
Train with RL
Rollouts are graded against your reward, and the model is trained until it clears your experts' baseline.
04
Evaluate honestly
Held-out real work, head-to-head with frontier models, with cost and latency reported alongside quality.
05
Deploy in your perimeter
VPC, on-prem or Carraggon cloud, behind your SSO, with audit on every call and full rollback.
06
Keep improving
Production feedback flows back into the reward, so the model gets better at your business each month.
Outcomes
Expert-level
quality on the specific task, not general benchmarks
10–100×
cheaper per task than routing everything to a frontier model
Weeks
from scoped task to a measurable, evaluated model
Yours
weights, data and evals stay under your control
Security & control
DSES is designed for regulated and sensitive environments first — everything the model learns from, and everything it produces, stays inside boundaries you set.
Your data stays yours
Training data is never pooled or reused across customers, and never used to improve anyone else's model.
Deploy where you must
Carraggon cloud, your VPC, or fully air-gapped on-prem for restricted workloads.
Auditable by design
Every prompt, rollout, tool call and model version is logged and reproducible.
Human in the loop where it counts
Confidence thresholds route the hard cases to your experts — and those decisions become training signal.
How is this different from fine-tuning?
Fine-tuning imitates examples. DSES trains against outcomes: the model attempts the real task in an environment, is graded on whether the result was right, and is optimised on that signal. That is what closes the last gap to expert level.
How much data do we need?
Less than most teams expect. A few hundred well-graded examples plus a working environment usually beats tens of thousands of unlabelled documents, because the reward carries the information.
Do we own the model?
Yes. Weights, evaluation sets and the reward definition belong to you, and can be exported or run in your own infrastructure.
What if a frontier model is already good enough?
Then we will tell you. Every engagement starts with a head-to-head evaluation; DSES is only worth building where the specialised model is measurably better, cheaper or faster on your work.
Pick the job your best people spend too much time on. We'll scope it, benchmark today's frontier models against it, and show you what a domain-specific expert system would do instead.
