01 / THROUGHPUT
340M
Events processed daily.
CONNEXT · APPLIED AI SYSTEMS · EST. 2024
Agents, retrieval and decision systems that run inside your stack, under your policy, measured against your numbers.
Five stacked layers: sources feed ingest, ingest feeds retrieval, retrieval feeds the agent runtime, and the runtime writes back to the systems your people already use.
Static claims, not live instrumentation: throughput 340 million events per day; median time to value 11 weeks; agents in production 1,180.
340M
Events processed daily.
99.98%
Pipeline uptime, trailing 12 months.
1,180
Agents in production, not in pilot.
11wk
Median time to a first production workflow.
AGGREGATE OF CONNEXT DEPLOYMENTS · 2025
CX-WEB-001 · REV 4
INTEGRATION SURFACE
Connext connects to SAP, Snowflake, Databricks, Salesforce, ServiceNow, Workday, NetSuite, Microsoft 365, SharePoint, Postgres, Kafka, S3, Okta, Azure OpenAI, Bedrock, Vertex AI, Jira and Twilio, and runs in financial services, insurance, manufacturing, logistics, healthcare payers and the public sector.
DRAG THE RAILS · ARROWS STEP
The model was never the hard part. The hard part is an answer becoming an action inside a system that already exists, owned by someone, audited by someone else.
Chart description: a pilot curve flattens after the first demo gains, while a production line steps up as grounding, evaluation, ownership and cost control are added.
Nobody agreed what a correct answer looked like before the build, so nobody can say whether it got better.
No policy sat between the model and the system of record, so legal stopped it a week before launch.
The build team moved on and nothing was instrumented, so failures arrived as complaints.
Token spend grew faster than the value it produced, and nobody held a ceiling.
They decide what we take on, and they are why the work survives its first incident.
NON-NEGOTIABLE · REV 4We ship into your stack, behind your auth, against your data. If it cannot run there, we say so in week one and stop.
Every answer carries its sources and every action carries its trace. If a claim cannot be attributed to a document, a record or a tool call, the system returns nothing.
Your data stays inside your boundary and you choose the model. Residency, retention and redaction are configuration, not a support ticket.
Every build runs the same five stations. Buy one or all of them; the interfaces between them do not change when you add the next.
CX-PL · 05 STATIONS · ONE CONVEYOR
Five stations in order: ingest, ground, reason, act, govern. Each lists three specifications and connects to the next along one conveyor line.
Diagram: seven source systems run into one vertical collector bus, which feeds a normalisation block in batch, change-data-capture or event mode, and leaves as a single record stream.
Connect the sources you already own and keep them current. Batch for history, change data capture for records, event streams for anything that moves hourly.
Diagram: a source document is split along a fan into five chunks, each embedded into a vector and keyword index shown as a lattice, and queried once across both a BM25 lane and a vector lane.
Chunk, embed and index your content so an answer traces back to the line it came from. Hybrid retrieval keeps keyword precision and adds semantic recall in one query.
Diagram: a task branches through six bounded depth columns of square nodes, converging on one answer node; a dashed lane leaves the tree below the confidence floor and runs straight to a deterministic terminal.
Agents plan against your tools, not the open web. Depth is bounded, every step is logged, and a deterministic path takes over below your confidence threshold.
Diagram: a gantry beam carries a carriage and arm over five write slots; the arm passes through an approval gate before it seats a record into the third slot, and a dashed lane returns to the first slot for rollback.
Writing back is where pilots stop and systems start. Every write is idempotent, reversible one action at a time, and gated by the approval your process already requires.
Diagram: a log table of records with several fields redacted, one row framed as audited; alongside, a stepped spend line climbing release by release toward a dashed per-team ceiling it never crosses.
Prompts, tool calls, retrieved sources and spend are logged per team. Every release is graded against a fixed suite before it reaches production.
Each ships as working software in your environment, with a spec an engineer can evaluate before anyone books a call.
5 LINE ITEMS · CX-S01 – CX-S05An agent gets one job, a fixed set of tools and a hard limit on what it may do without a human. It works tickets, quotes and claims end to end, then hands off at the edge of its mandate.
One index over the content your teams argue about: policies, contracts, manuals, tickets and the records behind them. Permissions are enforced at query time rather than filtered afterwards.
The processes that take days, cross three systems and currently live in a spreadsheet. Steps are durable, retries automatic, approvals stay with the people who own them, and any run can be replayed.
Ranked recommendations with a stated confidence and the evidence behind each one. You set the thresholds for auto-approve, review and reject, and move them without waiting for a release.
One control plane for every model your company uses, hosted or open weight. Route by task and cost, grade each release against a fixed suite, and give finance a per-team ceiling that holds.
Every stage ends with something you own, whether or not we continue to the next one.
SCHEDULE · CX-S01 – CX-S04Delivery runs as four stages over thirteen weeks. Weeks 0 to 2, discovery, ending with three systems mapped. Weeks 2 to 5, grounding, ending with one index live. Weeks 5 to 10, build, ending with one workflow in staging. Weeks 10 to 13, run, ending with a 99.9 percent SLA in force.
The scored shortlist and the system map, whether or not we continue.
A live index inside your perimeter and the eval set that grades it.
Source code, infrastructure as code and a staging deployment in your account.
A production system under SLA, with the runbook and the keys.
One team from week 0 to week 13. The engineers who scope it are the engineers who ship it.
The stations do not change from sector to sector. The vocabulary, the records and the regulator do.
6 PLATES · CX-IMG-0031 – 0036Onboarding files, KYC packs and credit memos checked against policy before a human signs.
First notice of loss triaged, documents extracted, claims routed with the reasoning attached.
Maintenance logs, SOPs and supplier quality data made answerable on the line, in the plant’s own language.
Shipment exceptions caught, rebooked and communicated before the customer calls to ask.
Prior authorisation and member correspondence drafted against policy, every citation on the record.
Casework and citizen correspondence handled inside an air-gapped boundary with a full audit trail.
ALL PLATES 16:10 · MIN 1600 × 1000
CX-WEB-001 · REV 4
Connext installs as a control plane inside your network. Your sources stay where they are, model calls leave only as the payload you approved, and every retrieval, tool call and write is logged on your side of the line.
Reference deployment diagram, read left to right. Outside your perimeter sit two things only: your data sources — document stores, databases, ticketing and event streams, reached by read-only connectors — and your model providers, whether a hosted API, a private cloud endpoint or open weights you host yourself. Inside your perimeter, five components run in dependency order. The Connext control plane owns connectors, scheduling and normalisation. The retrieval layer holds the hybrid index and its permission filter. The agent runtime plans, calls tools and falls back to a deterministic path. Policy and audit applies redaction and approval gates and keeps the prompt, tool and source log. Your system of record — ERP, CRM or ticketing — remains authoritative and receives idempotent, reversible writes. Exactly two connections cross the boundary: read-only ingest from your sources, and a single outbound inference call carrying only the payload you approved. The index, the audit log and every write stay inside.
Anonymised at each client’s request. Figures are read from the deployment’s own dashboards over a trailing twelve months, and references are available under NDA.
3 DOSSIERS · CX-IMG-0041 – 0043First notice of loss arrived as PDFs and voicemail, and handlers spent 40% of the day reading it.
€4.1M run-rate saved
Thirty years of maintenance logs and SOPs existed, but nobody on shift could find the right answer inside an hour.
$6.8M downtime avoided
Shipment exceptions were found by the customer before they were found by the network.
£2.6M revenue retained
Eight further deployments, each one sector and one metric: retail bank, minus 38% average handle time; property and casualty insurer, 61% auto-triage; automotive supplier, minus 22% scrap; third-party logistics operator, 3.4 times exceptions cleared; health payer, minus 52% touch time; utility, nine weeks to production; asset manager, 4,500 documents per day; city agency, air-gapped and live.
Slip 1 of 8RETAIL BANK · −38% AHT
P&C INSURER · 61% AUTO-TRIAGE
AUTO SUPPLIER · −22% SCRAP
3PL OPERATOR · 3.4× EXCEPTIONS
HEALTH PAYER · −52% TOUCH TIME
UTILITY · 9 WEEKS TO PRODUCTION
ASSET MANAGER · 4.5K DOCS/DAY
CITY AGENCY · AIR-GAPPED, LIVE
8 SLIPS · ANONYMISED · NDA ON FILE
CX-WEB-001 · REV 4
Your security team will ask these questions in week three. Publishing the answers saves a call and tells you what we will not bend on.
If an answer does not match your policy, say so in the first conversation. Some of it is configuration. Some of it means we are the wrong supplier, and week one is a better time to find out than week nine.
Deployment topology diagram, three states built from the same eight parts: sources, ingest, index, runtime, audit, keys, model and updates. In the VPC state the security boundary encloses sources, ingest, index, runtime, audit and keys; the model endpoint and the update channel sit outside it and cross the boundary. In the on-premises state the model moves inside the boundary and only the update channel remains outside. In the air-gapped state every part including the model sits inside a double boundary and the update channel is severed from it entirely.
BOUNDARY · YOUR CLOUD ACCOUNT We deploy into your cloud account. You hold the keys, the network policy and the bill.
BOUNDARY · YOUR DATA CENTRE Your hardware, your data centre, no outbound dependency at inference time.
BOUNDARY · NO ROUTE OUT Open-weight models on isolated infrastructure. Updates arrive as signed artefacts, by hand.
MOST AI PROJECTS DIE
BETWEEN A DEMO
AND A DEPENDENCY.
WE BUILD ONLY WHAT YOUR
BUSINESS CAN DEPEND ON.
THE OFFER
No slides. We look at one of your processes and tell you whether it is worth automating, including when it is not.
Throughput gauge, a static claim rather than live instrumentation: the platform carries 340 million events per day.
Tell us the process and the systems it touches. An engineer replies, and you will know within a week whether it is worth building.
The locator is an abstract orthogonal street grid with a single marker on the block holding our office at 35 Dunav Street, in Sofia's Oborishte district. It is a drawing, not a map, and carries no scale.