Buena + Qontext flagship

Clean context for agents that can’t afford to guess.

Context Surgeon turns scattered property-management data into living Markdown and VFS context, then keeps it current without overwriting the work humans already fixed.

See the architecture

Living VFS

agent-readable property memory

Fact Patch

refresh context without erasing edits

Agent Check

know when it is safe to act

incoming sourcecontradiction

Roof quote rejected

A new owner email changes what the agent should do next.

context-surgeon://sonnenallee-44

health 91

live compiler layer

Fact Patch

Generated markdown updates surgically while human notes remain untouched.

changed facts

+ TempSeal approvedcited

- quote pendingpatched

preserved: tenant prefers emailverified

sources

facts

conflicts

The context layer agents were missing

Stale context breaks agents.

Leases change. Vendors update terms. Owners override defaults. Agents keep acting on yesterday’s facts unless the context layer is rebuilt with intent.

Living Markdown + VFS

Messy property inputs become structured, agent-readable Markdown and virtual files.

Source-grounded facts

Every claim carries a quote, source span, confidence score, and conflict status.

Fact Patch preserves work

Generated context refreshes at fact level while human edits stay intact.

Hackathon sample-data run

Inazuma.co becomes a context repository, not a prompt dump.

The supplied dataset spans CRM, HR, enterprise mail, support, policy PDFs, IT tickets, GitHub, social data, and order PDFs. Context Surgeon compiles it into a file system, graph, review queue, and patchable update layer.

Raw records

153,997

Domains

VFS files

What the judges need to see

We start with raw company state, produce an inspectable VFS plus graph, preserve fact-level provenance, and involve humans only where ambiguity matters.

context-surgeon://inazuma.co/compile

Raw dataset -> VFS -> graph -> human review -> Fact Patch

sample dataset compiled

Raw company state

Top source domains by volume

12 domains

96,884

employees.json · resume_information.csv

CRM

14,951

customers.json · products.json · sales.json

Product Sentiment

13,510

product_sentiment.json

Enterprise Mail

11,928

emails.json

Inazuma Overflow

10,823

overflow.json

Collaboration

2,897

conversations.json

Entities

Edges

Provenance

fact span

Updates

Fact Patch

Static memory

3 files

Customers, employees, products, vendors, orders.

Procedural memory

1 file

Policy PDFs, rules, security, SDLC, leave policy.

Trajectory memory

3 files

Tickets, mail, GitHub, collaboration, support progress.

Generated repository files

source-linked

/company/inazuma/context.md

Company Operating Context

Dense executive context for AI agents and humans.

/company/inazuma/static/customers.md

Customers and Revenue

Static customer, product, sales, and order memory.

/company/inazuma/static/employees.md

Employees and Roles

Employee directory, HR relationships, resumes, ownership references.

/company/inazuma/procedures/policies.md

Policy Repository

Procedural company rules from policy PDFs.

/company/inazuma/trajectory/projects.md

Projects and Progress

Trajectory memory from mail, collaboration, GitHub, support, and IT tickets.

Maintained by Fact Patch

fact -> source record -> source file/path -> quote/span. New or changed records update the right files without destructive regeneration.

Starts with company data, not an agent.

Produces a virtual file system plus graph.

Preserves references inside files and back to source records.

Separates static, procedural, and trajectory knowledge.

Uses human review only where ambiguity matters.

Optimizes for long-term maintainability by humans and machines.

Recording mode

The full sample-data demo runs without sign-in.

checking accessSign in for live mode

Demo workbench

Operate on context before agents operate on the business.

Load source material, compile context, preserve a human note, ingest new evidence, apply a Fact Patch, and watch the agent action plan change.

phase: Archive loadedpublic preview

Recording path

Four clicks tell the full Fact Patch story.

2 min ready

Preserved

human note remains untouched

Patched

only stale generated facts change

Changed

agent plan flips to cited action

Audit timeline

Every context mutation leaves a trail.

preview state

Live intake available

Connect real systems after the sample-data loop.

The recording path stays focused on the provided demo data. Live mode can still sync Gmail, CRM, files, support, collaboration, work tracking, and manual uploads.

17 connectors

Buena hard problem

Signal vs noise before context.

Pioneer/Fastino classifies sources before extraction so irrelevant emails do not pollute `context.md` or downstream agent plans.

Partner technologies

Gemini

structured reasoning, conflict explanation, agent pre-flight

checkingstructured extraction

Tavily

external enrichment and verification for public/vendor context

checkingweb enrichment

Pioneer

schema-first source relevance and extraction pre-pass

checkingsource relevance

Sources

Facts

Open conflicts

Context health

40/100

Tenant complaint: roof leak still active

001_tenant_roof_leak.txt · 1c7d2bb4

signal · 94%2 facts

Signal classifier decision

Tenant reports active water ingress, status, and a response deadline for Sonnenallee 44.

classifier: pioneer-fastino-mock

From: Klara Hoffmann <klara.hoffmann@example.de>
To: property-manager@hausmail.de
Subject: Dachleck in Wohnung 4B weiterhin offen
Date: 20 Apr 2026

Objekt: Sonnenallee 44, 12045 Berlin.
The roof leak above unit 4B is still open.
Water marks expanded after the weekend rain and the tenant asks for a status update by 22 Apr 2026.
Please do not mark this as resolved until the ceiling is dry.

Extracted from this source

fact_roof_status_open_tenantfact_roof_deadline_status

Agent pre-flight delta

before patch

after patch

Persisted audit events

Sonnenallee 44 loaded

workspace.loaded

Public preview loaded the bundled property archive.

Sponsor-native architecture

Built around partner tech jobs, not logo stacking.

Gemini handles structured reasoning, Tavily enriches and verifies external evidence, Pioneer/Fastino supplies schema-first extraction hints, and the Qontext export makes the property memory inspectable by humans and agents.

Sources

emails, PDFs, CSVs, handover notes

Pioneer pre-pass

document kind + likely entities

Gemini extraction

typed facts + source quotes

Conflict engine

stale, missing, contradictory context

Fact Patch

minimal markdown diff with preserved human edits

Agent check

safe action plan with citations