DIH AI OS  ·  Executive Update

One system. One memory.
Every deal the firm touches.

The DIH AI OS unifies Atlas, Hub, and Email Triage on a single governed intelligence pipeline — turning every email and document the firm receives into permanent, permissioned, cited institutional memory. Keystone, and every domain agent it spins up, will run on this one unified memory layer.

11 June 2026

Strictly Private & Confidential  |  Internal — DIH AI Programme

From the last review

Three agreements set the direction. This update reports what was built on them.

01

Context first

The models need structured context — pipeline data, defined personas, curated entities — so the system stops guessing and starts knowing.

02

Governance first

Access controls, document protection, and audit logging form the foundation — trust built in before scale, not retrofitted after.

03

The AI OS is the critical path

One platform binding Atlas, Hub, and email into a single system — and, on your read, the seed of the firm's project pipeline. Design first, then build.

What follows is that design — broader and stronger than what we discussed in the room, documented decision-by-decision in a single system of record, and ready to build.

One unified AI OS

The next logical step after three proven apps: one institutional memory that feeds every app — and every agent after it.

Today, every system the firm runs reads the same mail and reaches its own conclusions. The unified AI OS inverts that: a single governed pipeline reads every source once, distills it into scored, permissioned facts, and lands them in one memory that every application — and, in time, every agent — draws on.

The architecture is fully documented — covering ingestion, entity resolution, confidentiality, data ownership, and the transition path for the current apps. Nothing here is a sketch — it is a ratified design with a phased build plan behind it.

And the memory compounds. Every deal screened, every model reviewed, every negotiation thread becomes queryable, cited evidence — the firm's accumulated judgment, available to every future deal team from day one.

Foundation The hardened M365 tenant — identity, classification, and source of every email and document. Unchanged.
Feeders One governed ingestion pipeline reads every source once and turns it into scored, permissioned facts.
Memory & Agents The institutional memory all apps share — and the substrate Keystone's agents will reason over.
Trust Entitlement, audit, and provenance enforced at every layer — built in from the first record.

The four layers of the DIH AI OS — each layer serves the one above it.

Atlas, Hub, Triage — same apps, stronger foundation

All three move onto the unified AI OS — re-seated, not replaced.

Today

Three independent AI pipelines

Atlas, Hub, and Email Triage share a database instance, but each runs its own independent AI pipeline. The same email is read three times, interpreted three ways, and remembered nowhere. The capability is proven; the intelligence is siloed.

M365 email · documents Atlas pipeline own ingestion + own AI Hub pipeline own ingestion + own AI Triage pipeline own ingestion + own AI Shared DB instance only
The Transition

One pipeline, three apps, one memory

The unified feeder reads each source once and serves all three apps from one governed memory. Each app keeps its job — Atlas the relationship graph, Hub the document home, Triage the inbox — and each switches over only when the unified pipeline proves more accurate than its own. The same memory then serves Keystone, the orchestrator, which spins up domain-specific agents — all reasoning over the one governed record.

M365 read once Unified Feeder one governed pipeline scored · permissioned One Memory Atlas Hub Triage Registers + briefs
Continuity The current apps keep running unchanged throughout the transition. Each one moves to the unified memory only after passing formal accuracy gates against its own results — users see no disruption, and nothing is switched off until its replacement is measurably better.

Follow a document through the pipeline

From the moment an email arrives to the moment it lands as governed, cited knowledge — automated end to end, audited at every step. This is how the memory gets built.

Press Play to watch emails travel the pipeline, or click any stage to inspect it.
Takeaway Every fact the firm acts on is captured once, scored, permissioned, and traceable to the exact email it came from — and the pipeline gets more accurate every day it runs.

Four properties no off-the-shelf tool offers

The moat is not the apps — it is the governed memory beneath them.

01

Ingest broadly, expose narrowly

Full email and document content is ingested, but every derived fact inherits the strictest permissions of its sources. No summary is ever readable more broadly than what it was built from — MNPI-safe by construction.

02

Accuracy gates

Each app moves onto the unified memory only when the new pipeline measurably beats the app's own results on adjudicated, side-by-side comparisons. Quality is gated, not assumed.

03

The feedback flywheel

Every human accept, reject, or correction in any app flows back into the memory and recalibrates it daily. The system the firm uses is the system the firm trains — a compounding asset competitors cannot copy.

04

Full provenance

Every fact carries its complete chain of evidence — which email, which version, which model, which human confirmed it. Audit, eDiscovery, and "why do we believe this?" are answerable in one query.

Takeaway The answer to "across every tower deal we've seen, what was region-specific vs. our standard model?" becomes an IC-grade, cited report — drawn from the firm's own accumulated judgment.

Agentic loops — how we accelerate the build

Autonomous build–verify–correct loops run many workstreams in parallel — the same step-change we are delivering for the firm is the one we use to build it.

How software was built

Manual coding

Engineers write every line by hand. Throughput is bound by headcount and hours; quality is bound by review capacity. One workstream per engineer, sequential by nature.

PaceBound by typing speed — weeks per feature.

How most teams build today

Agentic coding

Engineers prompt an AI agent task by task. The writing is faster, but every step still waits on a human instruction — one conversation, one thread of work at a time. The bottleneck moves from the keyboard to the prompt.

PaceBound by prompting — days per feature, one thread.

How we build

Agent-loop development

Autonomous loops plan, build, verify, and correct continuously. Engineers set direction and review decisions, not keystrokes — many workstreams run in parallel, each verifying its own output before a human ever sees it.

PaceParallel loops, continuous verification — a platform in weeks.

This is why the roadmap below is measured in weeks, not quarters. The design phase produced a decision-complete blueprint; agent loops now execute it in parallel while the team governs the decisions that matter.

Two phases to the full AI OS

Phase 1 is scoped, designed, and ready — the clock starts when development begins.

Phase 1 · AI OS Launch

The unified pipeline — and everything agreed

~3 weeks from development start

The feeder pipeline live end to end; Atlas, Hub, and Triage re-seated on the unified memory; the five registers populated and steward-curated; entitlement, audit, and the feedback loop running from day one.

Pipeline live
Apps re-seated
Registers populated
Launch
Phase 2 · Keystone + Subagents

The agents move in

Follows Phase 1 — on top of live memory

Keystone, the orchestrator, routes work to specialised subagents — deals & diligence, legal, investor relations, portfolio, and the firm's internal functions. Each agent reads the same governed memory and never exceeds the access of the person it serves. Phase 1 is what makes Phase 2 possible: agents are only as good as the memory they stand on.

System demonstration on live deal data

Two active mandates — Heirloom and Mesec — run end to end through the full pipeline.

Demonstration 1

Heirloom

Live data from ~February 2026

The full feeder pipeline run over the Heirloom correspondence and documents — watch real emails become scored, permissioned facts in the registers and surface in all three apps.

Demonstration 2

Mesec

Live data from ~November 2025

A deeper history: seven months of Mesec material through the same pipeline — demonstrating backfill over a longer window and a register that reads like the deal's institutional memory.

What the demonstration covers

Feeder data pipeline Triage Atlas Hub Deals register Entities register Suppliers register Investors register Internal Functions register

All live data available across the unified registers at the demonstration.

Takeaway The design is done, the method is proven, and the first launch is weeks away — judged on live results from the firm's own deals.
Open the DD Workbench

Design mockups — opens in a new tab.

DD Workbench: https://dih-demo.augmentgroup.ai/dih-dd-workbench.html