Built by operators from
For investment teams: a governed Strategy Unit researches and monitors each mandate around the clock, surfaces evidence-backed recommendations, and waits for your approval before anything moves.
Personalized portfolios, model sleeves, and house-view variants multiply faster than expert review capacity. The work that actually protects clients is exactly the work that doesn't scale by hiring: re-checking the thesis as markets move, documenting the why, and surfacing the exception before it becomes a surprise.
Mandate constraints, house views, and client preferences live in IPS PDFs and email threads, not in anything that can act on them.
A backtest run in March is forgotten by June. The reasoning behind a position rarely survives to the next review.
A clever chat answer is not a monitored mandate. Nothing accumulates; nothing is reproducible; nothing carries forward.
Every capital decision should arrive with its rationale, the data it used, and a record a reviewer or regulator can trust.
Citadel, Millennium, Jump, Tower, and Jane Street run dozens of semi-autonomous pods, each a small team built around a single mandate. What makes it work is the operating model: data, compute, implementation, risk, and audit are shared platform services, so a pod can launch, scale, constrain, or retire without rebuilding any of it.
ItoFlow reimagines the pod itself as a durable Strategy Unit. Give it a mandate and the capital to run, and an AI agent manages it perpetually on the same shared rails. That is what makes the model usable far beyond the mega-fund tier.
A Strategy Unit is a persistent, governed workspace for a single mandate: memory, research tools, an approval gate, and an audit trail. Here is the loop it runs, shown in the live product at every stage.
Describe what to own, what to avoid, and the risk posture to respect. ItoFlow compiles it into a typed strategy: triggers, eligible universe, constraints, and rules.
Ask in plain language: compare sleeves, test a signal, stress a risk posture. The unit writes the code, runs the backtest, and shows the assumptions behind every number.
One surface for state, holdings, review cadence, and portfolio context. Drift, failed runs, and stale data surface before they become operational surprises.
The unit packages a proposal with its rationale, evidence, risk context, and sizing impact. You approve, reject, or send it back. A person decides, and the choice is on the record.
This is not a one-off run. The unit keeps monitoring as markets move, accumulates evidence with every cycle, and sharpens its view of what is working over time.
We built the infrastructure to run them by the thousand. Each mandate gets its own durable, isolated agent that runs for months without losing context, and a control plane built to fan out to thousands in parallel. An entire multi-strategy operation, run as software.
Each strategy runs as a durable, checkpointed workflow that resumes after any crash or restart. And because an agent's context fills up fast, every Strategy Unit keeps a long-term memory it rehydrates each cycle, recalling the signals it has tested and the calls a reviewer rejected. It runs for months and never starts cold.
Every mandate runs in its own isolated Firecracker sandbox. Agents fan out in parallel, and a failure in one never touches another. The control plane is built to orchestrate thousands at once.
Mandates compile to a versioned specification: signals, triggers, constraints, and decisions as validated schemas. Strategies are diffable, testable, and reproducible months later, never brittle prose.
When an agent reports a number, it links back to the signal, the information coefficient, and the backtest behind it. Claims carry their evidence, so a reviewer audits the case, not a black box.
Before a decision reaches you, a critic agent re-runs the backtests and stress-tests the risk, then archives its critique alongside the decision for the record.
Agents act only through a sanctioned tool registry with policy checkpoints. Every action is permissioned, logged, and revocable, so an agent can do its work without ever holding the keys.
ItoFlow is built for regulated mandates. The AI does the research and assembles the case; a person makes every capital decision; and the record is complete enough to defend.
Recommendations never reach the market on their own. Capital decisions sit behind an explicit approval gate: approve, reject, or send back.
Each proposal carries the data it used, the code it ran, the assumptions it tested, and the risk context, so a reviewer evaluates the case, not a black box.
Every decision, accepted or rejected, is logged with its reasoning and inputs. Reconstruct any recommendation, any day, for review or audit.
Your data and your mandates are never used to train models. Workspaces are isolated per mandate, and you choose the deployment mode.
Start narrow, where operational coverage matters and expert time is scarce. Expand as the record earns it.
Keep many client-specific mandates continuously researched and monitored, not just the top accounts.
Run model portfolios and sleeves with documented reasoning and a consistent review cadence.
Turn ad-hoc research into review-ready recommendation packets your IC can actually act on.
Run a new strategy in shadow alongside the live book before any capital is committed.
Surface mandate drift, stale data, and unusual behavior before the next scheduled review.
Give compliance a reproducible trail behind every recommendation and approval.
Prove coverage, control, and review quality on a focused road test for the ideas your team hasn't had the capacity to run.
Pick one place coverage is constrained: a sleeve, a mandate set, or a research backlog.
Wire in the data access, constraints, and review cadence the workflow needs.
Let Strategy Units research, monitor, and recommend, with no capital at risk.
Compare the record, promote what earns it, and decide where ItoFlow sits in governance.
Head of MFT at Tower Research, ex-J.P. Morgan. A decade building hedge-fund research and trading infrastructure. IIT Kharagpur.
Nine years building massive-scale infrastructure at Google and Apple. Products serving 100M+ users. IIT Kharagpur.
Mission-critical systems at Microsoft and Azure DNS. Expert in distributed AI agent orchestration. IIT Kharagpur.
Stanford PhD. Former J.P. Morgan London Head of Quant Research. Senior AI research director.
Cover more mandates, move faster, and keep a person on every decision. Proven on your own book in a single workflow.