Working Agents · For Regulated Funds

We ran agentic AI inside a leading hedge fund. Now we build working agents inside yours.

The full factory inside your perimeter in two weeksturn-key if you have no infrastructure at all. Your people build workflows in plain language; when a system calls for engineers, we co-develop it, guaranteed.

Claim your first seat
iOperators firsttwo years running agentic AI inside a leading hedge fund
iiUnder real stakesNAV, administrator, auditor — the discipline ran under fund-auditor scrutiny
iiiDelivered bya 15-year enterprise delivery organisation, engineers embedded
01 The backlog

The systems your team always wanted, finally built.

Every fund keeps a quiet backlog: tools the desk asked for years ago, still waiting behind whatever internal engineering had to do first. We put frontier intelligence beside the people who own those processes, co-creating until the backlog moves.

aCapital formation

LP pipelines researched and kept warm, dossiers ready before every meeting, DDQs and data-room requests handled overnight. Partners show up to close.

bResearch & deal flow

Every filing, transcript, and data room in your universe read overnight and scored against your thesis. First drafts arrive with citations; analysts start at the interesting part.

cQuant & data ops

Vendor feeds onboarded, validated, and watched. Backtests run and written up, anomalies investigated before they touch a signal. Quants stay on alpha.

dMarkets & treasury

Breaks, fails, margin calls, and covenants worked overnight across admins and prime brokers. Exceptions reach a person with the evidence attached.

eCompliance reporting

Form PF, Annex IV, and 13F assembled from the book, every number traceable to source. Marketing review and personal-trading surveillance run continuously; sign-off stays with your CCO.

fInvestor relations

Quarterly letters drafted in your voice from your numbers, LP queries answered same-day, KYC refresh queued for sign-off. The relationship time stays with the partners.

Built for multifamily offices, private equity, and hedge funds, front office to back. What ships is the system your team designs — and the hours come back to the people whose judgment is the business.

02 The engagement

Experience it first. Expand by the seat. Co-develop what's heavy.

01 · Experience

On us

one seat · live in 2 weeks

The full harness, deployed for one principal. Brief the Chief of Staff, build your first workflows yourself — feel it firsthand. Converts to your first principal seat.

02 · Expand

from £1,500

per seat · per month

Principal seat £3,500 — the chief of staff, priority support. Team seat £1,500 — process owners self-developing in the factory. Operation, evals, and upgrades included; the brain compounds across seats.

Stack
Claude-first, with private open-weight models for role-specific sub-agents. Your cloud, inside your security perimeter.
No infrastructure?
Turn-key deployment: we provision a dedicated single-tenant environment in your name — your account, your keys, our runbooks. Infrastructure billed at provider cost, itemised on the Ledger.
Model spend
Billed at provider cost, itemised per task in your Ledger.
Ownership
Your workflows and data are yours outright. The Genba harness carries a perpetual license for the deployed version, so it keeps running with or without us. What your teams build in the seats — the capabilities and the record — compounds as your asset, which is what makes leaving safe.
03 Why Genba
現場 — Genba

Genba (現場) is where the work actually happens: the desk, the book, the close. A craftsman is judged by what still runs after he leaves.

Most enterprise AI is sold from a slide and dies on contact with a real operation. Our engineers embed where the work happens and stay until an agent runs there in production.

04 The factory

We install a software factory. Your people run the line.

Your first seat puts the whole harness in place — gateway, ledger, evals, company brain. The first workflow is the first thing off the line, and that's how a backlog actually clears: the cost of building collapses, and it stays collapsed.

05 How it runs

Engineered to the standard a regulated fund demands. Every action is gated and recorded; people hold the decisions that carry consequences.

YOUR INFRASTRUCTURE — ON-PREM / PRIVATE VPC DATA RESIDENCY BOUNDARY OUTSIDE consequential actions → a named approver signs, or it doesn't happen YOUR PEOPLE via Slack · Teams · email direct sandbox — holds no credentials AGENT drafts · checks · proposes request scoped grant GATEWAY your policy · expiring credentials every grant recorded YOUR SYSTEMS OMS · admin portal · data · email PRIVATE MODELS open-weight · your hardware EGRESS SCREEN FRONTIER MODELS only what needs them anonymise or block — raw data never crosses LEDGER every action, inference and grant — one record, queryable at audit
  • Sandboxed and secret-free. Agents hold no credentials. The Gateway grants scoped, policy-checked access and escalates to a person when the stakes require it.
  • Audited to the token. The Ledger ties every action, inference, and dollar of model spend to the task that caused it. An audit becomes a query.
  • Humans hold consequence. Anything with stakes routes to a person, with the agent's reasoning visible enough to challenge.
  • No shadow automations. Every agent-built capability is registered, versioned, and held to evaluation thresholds. Query the registry: what exists, who owns it, what it can access, how it performs.
  • Watched from outside. Sentinels the agents cannot see stand watch. Security reads every log stream for unapproved behaviour — intentional or not — with freeze and kill-switch authority. Compliance screens every outbound flow, anonymising PII and blocking flagged content. Cost watches the Ledger and freezes budget runaways. All escalate to human triage.

The company brain is the institutional memory every agent reads from and writes back to, with provenance built in.

SUBSTRATE KNOWLEDGE GRAPH ontology of institutional knowledge structures EMBEDDINGS universal · encodes every data type encodes CORPUS every format the business runs on — text · image · video · documents ACCESS · PROVENANCE COMPANY-WIDE pull request → DIVISION pull request → WORKING MEMORY THIS AGENT personal history · operating memory read & write freely Every change is a Git commit. Shared tiers are gated by pull request and review; an agent's working memory is its own. All knowledge carries its origin.

Each finished task leaves the firm knowing a little more, and the knowledge stays when people move on. The agents are outputs; the asset is the institutional intelligence you accumulate — owned by you, exported on exit, appreciating as models commoditise.

A Chief of Staff — itself an agent — is briefed through the front doors your people already use: Slack, Teams, email. It works a shared task database in parallel and routes every task to the right mind for the job — role-specific agents on open-weight models you own, or the frontier when the work demands it. Nothing consequential leaves the system without passing the approval gate, and every send lands in the audit log. Your data stays inside the perimeter.

YOUR PRINCIPAL Slack · Teams · email directs, in plain language one answer, with evidence CHIEF OF STAFF assigns work · carries context · reports back oversight — assign · query · review INVESTMENT team + agents RESEARCH team + agents TRADING team + agents RISK team + agents OPERATIONS team + agents FINANCE team + agents COMPLIANCE team + agents IR team + agents YOUR SYSTEMS & DATA one governed context, carried across every division

The open-weight models are yours, deployed inside your perimeter and improving over time — RLVR fine-tuning and eval-driven prompt optimisation, per agentic role. The intelligence your operation accumulates is an asset you own outright, immune to vendor lock-in. The same gateway, ledger, and sentinels govern every route, so a regulator sees identical controls whichever model served the request. On the outbound path, the compliance sentinel screens every flow with locally hosted models of its own: PII is anonymised before anything crosses the boundary, and anything compliance-flagged is blocked and escalated to a person.

Research, end to end — a multi-agent harness inside your perimeter. One evening, one analyst, one structured note by morning.

18:20Your team

An analyst briefs the harness: the target, the thesis memo, the questions that actually matter.

18:25Research agents

The harness fans out — one agent reads the data room and filings, one runs the quant work.

overnightPrivate models

The heavy reading runs on open-weight models on your hardware. The data room never leaves.

07:30Research agents

One structured note: scored thesis map, quant appendix, contradictions flagged.

08:15Your team

The team challenges it in plain language; answers come back with sources attached.

Ledger

Documents read, models used, the lineage of every number: recorded per task.

The same pattern runs trading (signal to settlement, gated and booked) and investor relations (a 240-question DDQ returned the same day, signed by people).

06 Two models

Two models. One control plane.

For principals: a chief of staff. Arrive Monday to find the night's work done — the investor letter drafted, the overnight moves reconciled, the follow-ups queued for a yes or a no. Every engagement starts here: one seat, one principal, the fastest way to feel what delegating to an agent is like.

For the divisions: teammate agents. Inside each team, working the routine alongside the people who own it — fails and breaks worked before the desk sits down, filings read overnight, DDQs drafted from the approved library. Division-scoped access, the same controls throughout.

07 Built on Claude

We build inside Anthropic's ecosystem, on the frontier of what Claude can safely do in production.

08 The architects

You work directly with the people who build it.

Makoto Tominaga

Production AI · Funds

Two years running production agentic AI inside a leading hedge fund. Twenty-five years across kernel, security, and AI: AuthenTec (acquired by Apple), Validity Sensors (acquired by Synaptics), founder of Credify (BEENEXT and DEEPCORE-backed).

Alex Norris

Investments · Quant

Head of Investments at the same fund, risen from quantitative researcher to Director. MSc in Risk Management & Financial Engineering from Imperial College London; first-in-cohort MSc in Data Science at Bath. He has run the research the agents now serve.

Steve Lim

Engineering · Fund Systems

Head of Engineering at the same fund. The production and data infrastructure behind a decade of award-winning performance is built and operated by his team.

Toan Nguyen

Delivery · Engineering

CEO of Trustify Technology (Vietnam). Fifteen years of enterprise IT delivery at Nortel, SingTel, Ruckus Wireless and Absolute. Runs an established engineering organisation with managed-services discipline and Southeast Asian cost economics.

09 Start

Tell us what's sat on your backlog the longest.

Claim your first seat

Or reach the architects directly:

makoto@genbalabs.com toan@genbalabs.com