Fragmented institutional knowledge
Pricing logic, metric definitions, schema conventions and operational rules lived in people's heads, scattered chat threads and undocumented notebooks. Every new request meant reconstructing context from scratch.
CASE STUDY
An AI intelligence operating system for a specialist business.
A unified platform that turns scattered institutional knowledge, fragmented analytical tooling and recurring operational work into a single, agent-operable system. Designed so a small team can move at the speed of a much larger one — and so the knowledge built every day is the same knowledge available tomorrow.
- Engagement
- End-to-end platform build
- Team size
- Small, high-trust analytics team
- Surface area
- Knowledge, tooling, workflows, reporting
- Outcome
- Order-of-magnitude throughput gain
THE CHALLENGE
Operational knowledge and tooling that lived everywhere except in one place.
Inconsistent analytical tooling
Hundreds of overlapping SQL files, ad-hoc notebooks, duplicated patterns and hardcoded credentials. No registry, no shared runtime, no reliable way for a person or an agent to discover and reuse what already existed.
Manual, repetitive workflows
Recurring work — performance reviews, monthly reports, change evaluations, data quality checks — followed similar steps each time but was rebuilt by hand on every execution. Slow, error-prone and impossible to scale.
No feedback loop
Insights from completed analyses stayed locked inside individual folders. Validated findings, learned edge cases and new domain understanding never flowed back into the team's shared baseline.
BEFORE / AFTER
From scattered fragments to a single operating surface.
Before
Fragments, everywhere
Tribal memorySlack threadsLoose notebooksHardcoded SQLForgotten notesOne-off scriptsPersonal docsAd-hoc CSVs
After
One operating surface
- Knowledge in one navigable library
- Tools with a standard contract
- Workflows that any agent can run
- History that compounds, not decays
- Output contracts everyone trusts
THE SOLUTION
One system. Knowledge, tooling and workflows in the same place — built to be operated by an agent.
The platform is a structured workspace where business context, reusable analytical capabilities and codified workflows live side by side. It is designed so that an AI agent — operating inside any modern coding or analysis environment — can autonomously navigate context, select the right tools, execute the workflow, apply domain-specific business rules and produce stakeholder-ready outputs.
It is deliberately lightweight. There is no separate application server, no database backing the knowledge base, no orchestration platform to maintain. The repository is the system. That makes it zero-infrastructure to operate, version-controlled by default, and easy to extend without touching what already works.
ARCHITECTURE
Five interconnected layers, all navigable by a person or an agent.
Person or agent
Operating in a coding or analysis environment
Layer 01
Workspace
where the work happens
History
for reuse
Layer 02
Knowledge
what the business knows
Intake
for capture
Layer 03
Tools
what gets executed
Runtime
the foundation
Playbooks
Codified workflows that compose the layers above
THE LAYERS
What sits inside each layer of the system.
Layer01
Knowledge Layer
Compiled business context
A curated, interlinked library of Markdown pages that encodes the business domain — fundamentals, architecture, history, vertical context, validated metric definitions, schema guides and team conventions. Knowledge becomes navigable, version-controlled and agent-readable.
Layer02
Workspace Layer
Structured operational execution
Every piece of work flows through a defined lifecycle — classify, load context, check past work, plan, execute, apply business logic, deliver. Each request lives in its own folder with a complete analytical trail attached.
Layer03
Tool Layer
Reusable analytical capabilities
A registry of purpose-built tools and templates with a standard contract: documented inputs, outputs, dependencies, side effects and smoke tests. A shared runtime provides connections, parameterised queries, storage access and standardised output bundles.
Layer04
Playbook Layer
Codified workflows
Recurring procedures — review flows, reporting cycles, optimisation routines, evaluations, experiments — encoded as repeatable, agent-executable sequences. Production-learned edge cases are captured directly inside the playbook so the same bug never happens twice.
Layer05
Intake Layer
Structured knowledge capture
A deliberate process for closing knowledge gaps: targeted questionnaires, freeform expert input, and structured synthesis into wiki pages. Every gap exposed by a request can trigger an intake session that improves future work.
REQUEST LIFECYCLE
Every piece of work follows the same seven-step path.
Codifying the lifecycle means the right context is loaded, the right tools are used and the right rules are applied — every time, by anyone, including an agent.
- Step 01ClassifyTriage the request against defined archetypes.
- Step 02Load contextPull relevant knowledge pages automatically.
- Step 03Check past workFind and reuse comparable prior analyses.
- Step 04PlanCapture question, sources and approach.
- Step 05ExecuteRun registered tools and templates.
- Step 06Apply rulesValidate against critical business logic.
- Step 07DeliverFormat for the audience and archive the trail.
Request receivedOutput delivered
IN PRACTICE
What the system actually does, day to day.
Three representative scenarios. Each one would previously have absorbed an analyst's attention for hours. Each is now a structured run through the system that consistently applies the right rules and produces output a stakeholder can act on.
Example 01
Recurring performance review
A stakeholder asks how a key account is performing and whether anything should change.
- 01Open a structured request folder for the work.
- 02Load the relevant business fundamentals, vertical context and validated metric definitions.
- 03Run profile, performance and quality templates from the registered tool layer.
- 04Invoke the optimiser to evaluate the most material decision points.
- 05Synthesise findings into a stakeholder-ready note with explicit recommendations.
Outcome · Time-to-answer drops from hours to minutes, and every relevant business rule is applied automatically.
Example 02
Periodic management report
Leadership requests the monthly performance pack across the most important accounts.
- 01Identify the in-scope accounts using a single canonical query.
- 02Pull current period, previous period and prior-year metrics from the shared template library.
- 03Apply reference-aware delta logic to handle near-zero baselines and new accounts gracefully.
- 04Flag data-quality concerns inline rather than burying them in commentary.
- 05Produce a structured report with headline metrics, detail, per-account highlights and methodology notes.
Outcome · The same process runs every period with the same fidelity. Lessons from earlier runs are encoded in the playbook, so prior bugs never recur.
Example 03
Commercial scenario analysis
A commercial team asks what to expect from a prospective deal under a range of structural choices.
- 01Load deal-structuring principles and the relevant vertical context.
- 02Pull comparable benchmarks using shared templates with the correct exclusions applied.
- 03Run a scenario projection across a grid of operational and commercial assumptions.
- 04If a payload is available, identify which fields would be most informative as new dimensions.
- 05Produce a benchmark table, expected-performance summary and commercial advice tagged for the right audience.
Outcome · Sales conversations are supported by consistent, defensible analysis grounded in real precedent rather than ad-hoc spreadsheets.
RESULTS & IMPACT
What changes when the system is in place.
Throughput up by an order of magnitude
Work that previously absorbed hours of an analyst's time — locating context, writing SQL, recalling business rules, formatting output — now completes in minutes. The team handles a far higher volume of requests without adding headcount.
Consistency, structurally enforced
Verified metric definitions, standardised templates and encoded business rules eliminate an entire class of bugs. The output is correct by construction, not by recall.
Knowledge no longer walks out the door
Business context, analytical procedures and operational learnings live in version-controlled files. New people — and new agents — reach productive capability by reading the system, not by shadowing a colleague for months.
Tooling reuse instead of reinvention
Every tool, template and helper is registered, documented and consistently invoked. Before anyone writes new SQL or builds a new pipeline, the registry tells them what already exists.
Production experience, encoded
Playbooks document the exact bugs that occurred on first live runs. Tools maintain persistent state across runs and learn from past evaluations. Golden test cases pin outputs against known-good results.
THE COMPOUNDING LOOP
Every request makes the next one easier.
Insights, edge cases and validated procedures flow back into the system rather than sitting in private folders. The platform gets sharper with use, not noisier.
Step 01
Request
A new question arrives
Step 02
Execute
Run with current knowledge
Step 03
Capture
New insights and edge cases
Step 04
Encode
Into wiki, tools, playbooks
loops back
DESIGN PRINCIPLES
Decisions that hold the architecture together.
Repo-native, not platform-native
The repository is the system. No separate application server, no orchestration platform, no vector store to maintain. Zero infrastructure to operate, version-controlled by default, trivially portable.
Markdown-first knowledge encoding
Business knowledge lives in plain Markdown with light frontmatter. Simultaneously human-readable in any editor, agent-readable by any model, and fully version-controlled.
Agent-readable by default
Every structural decision — file naming, folder layout, index files, embedded instructions, scoped rules, standardised tool contracts — is designed to help a model navigate, retrieve context and reason effectively.
Engine + interface separation
Every tool separates its pipeline logic from its command interface. Tools can be invoked by an agent, by a person, or composed into larger workflows — without any code changes.
Standard output contracts
Every tool emits a human-readable summary, a machine-readable summary and a detailed artifacts directory. Downstream consumers — people, scripts or other tools — can rely on a consistent interface.
Incremental migration
Legacy tooling is not copied wholesale. Each capability is migrated, rewritten against the shared runtime, validated against live data, hardened from production experience and pinned with golden tests.
WHERE THIS GOES NEXT
Natural extensions once the foundation is in place.
Once the core is operational, the same architecture extends naturally into proactive monitoring, automated knowledge maintenance and richer cross-portfolio intelligence.
Automated monitoring and alerting
Shift the team from reactive to proactive: scheduled quality checks across all active accounts, with anomalies surfaced before stakeholders ask.
Self-updating knowledge base
Completed work proposes draft updates to the wiki — captured as reviewable changes — so the knowledge layer keeps pace with what the team learns.
Natural language request intake
Front-end integrations parse incoming requests, classify them, create the workspace and either execute simple work autonomously or prepare a structured brief.
Cross-portfolio analytics
A portfolio view layered on top of the per-account tooling: cohort benchmarking, trend monitoring, and early warning of systematic shifts across the business.
Counterfactual impact estimation
Generalise the evaluation pattern into a standalone capability that answers "what would have happened without this change?" for any operational decision.
External data enrichment
Integrate market benchmarks, seasonality indices and macro indicators to improve model inputs and enable richer commercial analysis.
A SYSTEM LIKE THIS, FOR YOU
Scattered context. Repetitive analysis. Knowledge stuck in people's heads. We build the system that fixes all three.
Every engagement starts the same way: a structured assessment of where the highest-leverage opportunities sit in your operations, and a credible plan for acting on them. No pitch. No generic roadmap. No commitment beyond the conversation.