Foundation models
Large language models trained for long-horizon reasoning, reliable tool use and long context.
- Structured tool calling
- Long-context retrieval and memory
- Sizes for cloud and private deployment
Large language models · AI agents
Models that reason. Agents that finish the job.
SuperAI builds the next generation of large language models and the agents that run on them. Our agents plan multi-step work, use your software the way a person would, and check their results before they report back.
agent run · example
complete
task Reconcile September invoices against the ledger and flag anything that doesn't match.
- planPull September invoices and ledger entries, match them by vendor and amount, then recheck whatever is left over.
- callinvoices.list(month="2026-09")→ 412 invoices
- callledger.query(account="AP", month="2026-09")→ 409 entries
- callmatch(key=["vendor", "amount"])→ 406 matched, 6 open
- checkLook up the 6 open items by invoice number→ 3 were posted in October; 3 are real mismatches
- doneFlagged 3 mismatches for review, with the source rows attached.
6 steps · 4 tool calls
00:41
What we build
One stack, from the model to the agent at work.
We train our own models and build the runtime and evaluation around them, so an agent's behavior can be traced from its weights to its last tool call.
Agent runtime
The layer that turns a model into a worker: planning, tool access, permissions and a full record of every action.
- Connectors to your APIs and data
- Approval steps you define
- Replayable run logs
Evals and safety
Task suites that measure whether an agent got the job right, not only whether its answer sounds right.
- Task-level success scoring
- Regression checks on every release
- Red-teaming for tool misuse
How an agent works
Every run follows the same four steps.
The loop repeats until the task is done or the agent reaches a limit you set. Each step is written to the run log.
-
Plan
The agent reads the task, splits it into steps and decides which tools each step needs.
-
Act
It calls your tools and APIs within the permissions you granted, and records each input and output.
-
Check
Before it reports, it verifies the result against the source data and retries or escalates what doesn't hold up.
-
Report
You get the outcome, the evidence behind it and anything that still needs a human decision.
How we work
Built to be trusted with real work.
Checked before it's reported
An agent's answer is only as good as the evidence behind it. Our agents verify their own results before handing them over.
Every action on the record
Each plan, tool call and result is logged, so any run can be replayed and audited step by step.
People set the limits
You decide which tools an agent can use, what it can change and which steps need your approval.
Measured on outcomes
We judge models by whether the task was completed correctly, and we publish how we measure it.
Research
What we're working on.
Long-horizon reasoning
Keeping a model on track across tasks that take dozens of steps and hours of work.
Tool use and computer use
Teaching models to operate APIs, browsers and desktop software with the precision of a careful operator.
Memory
Letting agents carry what they learned on one task into the next, without carrying over mistakes.
Agent evaluation
Benchmarks built from real workflows, scored on whether the work was done right.
Careers
Join us early.
We're a small team building models and agents from the ground up. If you want to work on hard problems with a lot of ownership, we'd like to hear from you.
- Research Scientist, ReasoningResearch
- Machine Learning Engineer, TrainingEngineering
- Software Engineer, Agent RuntimeEngineering
- Research Engineer, EvaluationResearch
Contact
Bring us the work you want an agent to do.
Tell us about the task, the tools involved and what a good result looks like. We'll get back to you within two business days.
zwang3286@gatech.edu