Domain intelligence for better models. Agentic AI for real workflows.
Zstate builds the training infrastructure that makes models smarter, and the AI-native systems that put them to work across domains.
Coding dataset
258k
Real software engineering tasks with full agent
trajectories, tool calls, and human acceptance signals.
Nothing synthetic.
3.7M
Trajectory
Steps
Steps
130k
Reward
Signals
Signals
1.7M
Tool
Interactions
Interactions
Explore →
Healthcare dataset
Explore → 5M+
Prescription digitisation, diagnostic reasoning,
radiology, pathology, and drug grounding.
Domain experts
5000+
Specialists across software engineering, healthcare &
finance
What we work on
Domain Intelligence
Training infrastructure for models: RL environments,
RLHF and SFT datasets, reward signals, and evaluations.
Agentic AI
AI-native systems that turn domain workflows into
production software, with automation, orchestration, and
human review where needed.
Generic annotation fails
where domain judgment matters
Generic annotation
- Labels created by people who have never done the work
- Tasks reduced to shallow fragments with little workflow context
- Benchmarks that reward looking right, not being right
- Little connection between training data and production behaviour
Zstate approach
- Tasks, rubrics, and reward signals designed with domain experts
- RL environments and evaluation loops grounded in real workflows
- Training infrastructure built with downstream system behaviour in mind
- Signals that help models learn judgment, not just pattern matching
Domain Intelligence and Agentic AI
01
Primary
Domain Intelligence
Expert-led data, evaluation, and feedback systems that keep specialist context inside the model-development loop.
- RLHF preference data & reward model training
- SFT instruction datasets from domain experts
- Red-teaming & adversarial evaluation
- Clinical NLP, diagnostic Q&A, EHR abstraction
- Earnings analysis, risk data, compliance evaluation
- Medical coding & ICD abstraction
02
Powered by the same expertise
Agentic AI
Production-grade agentic systems across software engineering, healthcare, and finance, built with domain context carried through design, evaluation, and deployment.
- End-to-end agentic system design & build
- Multi-agent pipelines & workflow automation
- AI-native architecture, not retrofitted legacy code
- From prototype to scalable production deployment
- Compliance-aware engineering for regulated industries
Four areas. Genuine depth in each.
01 - Coding 03 - Finance 04 - Agentic AI
Software engineering data
02 - HealthcareHealthcare records & reasoning
Agentic systems
Complete agent trajectories across 258k real-world software engineering problems with reasoning traces, tool calls, code edits, and explicit user acceptance signals. Three core layers: Task, Trajectory, and Reward datasets.
- 258k real engineering tasks. Complete agent trajectories across real-world software engineering problems with reasoning traces, tool calls, code edits, and explicit user acceptance signals. Nothing synthetic.
- Three derived datasets. Task dataset (258k cleaned prompts), Trajectory dataset (3.7M full agent traces), and Reward dataset (130k acceptance signals).
- Beyond SWE-Bench. Full lifecycle: reasoning → tool calls (6-7 per task across 22 tools) → code edits → human acceptance. Real production tasks, not curated benchmarks.
5M+ connected healthcare records covering prescription digitisation, diagnostic reasoning, radiology, pathology, and drug grounding mapped to symptoms, diseases, and side effects.
- 5M+ connected healthcare records. Prescription digitisation, diagnostic reasoning, radiology, and pathology report interpretation in one corpus.
- Prescription, diagnostic, and report workflows. Extraction and interpretation tasks spanning prescriptions, clinical reasoning, radiology, and pathology.
- Drug data that completes the corpus. Drug layer tied to symptoms, diseases, and side effects - the grounding context for medical AI training and evaluation.
Preference data and SFT datasets for earnings reports, 10-K filings, risk model assessment, regulatory compliance, fraud detection, and trade rationale evaluation.
- Earnings & analyst evaluation. Preference data and SFT datasets over earnings reports, 10-K filings, and sell-side research. Evaluated by credentialed analysts.
- Risk & compliance data. Training and evaluation data for risk modelling, regulatory tasks, and stress testing. Reviewed by risk professionals.
- Fraud detection & trade rationale. Expert-annotated datasets for fraud detection, trade rationale evaluation, and financial reasoning benchmarks.
Multi-agent system design, custom RL environments, and production deployment with guardrails, observability, human-in-the-loop checkpoints, and scalable infrastructure.
- Agentic system design. Multi-agent architectures with tool use, memory, orchestration, and handoff logic for long-horizon workflows.
- RL environment engineering. Custom RL environments that simulate real expert decision workflows. High-signal training data and meaningful evaluations.
- Production deployment & ops. Prototype to production with guardrails, observability, human-in-the-loop checkpoints, and reliable infrastructure at scale.
How we work
01
Scope the workflow
We define the task, decision points, quality bar, and
operational constraints.
02
Design the signal
We shape the right combination of datasets, reward signals,
evaluations, or RL environments.
03
Deliver in loops
Work is delivered in structured iterations with QA,
feedback, and calibration built in.
04
Deploy or extend
If the workflow needs to run in production, we build the
agentic system around it.
Some teams start with training infrastructure. Others start with the workflow. We support both.
Why Zstate
Domain experts, not generic labor
The core signal comes from people who understand the
work, not from low-context annotation.
Real workflows, not shallow proxies
We build around actual decision processes, including
tasks, rubrics, evaluations, and RL environments.
Training and production under one roof
The same company that shapes model signals can also
build the system that uses them.
Depth where judgment matters
We focus on software engineering, healthcare, and
finance, where expertise changes model quality.