The plant
already knows.
Nobody can find it
in time.
A refinery, a power station or a factory already owns the answer to almost every operational question it has. The answer is buried across drawings, procedures, maintenance history and field notes, in a dozen systems that do not talk to each other.
Kairos reads all of it into one system that answers questions in plain language and shows the source document behind every answer. When the evidence is thin on a safety question, it refuses rather than guesses.
The EQ-101 seal spec became FSL-2240B on 15 January 2025. A repair four months later still fitted the superseded part.
Evidence
- FP-SB-2025-04OEM15 Jan 25
- WO-2025-0631Site record15 May 25
- Vibration note, unloggedUnverified15 Jan 26
This is not a filing problem.
It is a safety problem.
A large plant runs on documents written over forty years by people who have mostly left. When a technician cannot find the right one in the ten minutes they have, they do the job from memory, either their own or someone else's. That is where incidents start.
35%
of the working day
Time people in asset-heavy industries spend looking for information, chasing someone for an answer, or rebuilding a document that already exists somewhere.
McKinsey, 2024
7 to 12
disconnected systems
The number of separate places one large plant keeps its drawings, work orders, procedures, inspection records and regulatory filings.
NASSCOM and EY
18 to 22%
of unplanned downtime
Outages caused by maintenance decisions made without the equipment's full history in front of the person deciding.
BIS Research
25%
of experienced engineers
The share of India's industrial engineers and operators who retire within the decade. Whatever they never wrote down leaves with them.
Industry estimate
Search alone cannot fix this, because the most dangerous gaps are the ones nobody knows to search for. A technician who has never heard that this pump failed this way before does not think to go looking. So Kairos does not wait to be asked.
One pump, one work order,
start to finish.
Everything below happens without anyone typing a search. Follow it once and the rest of this page explains itself.
- 1The trigger
A work order is raised on pump P-101
A technician is assigned to a seal failure. Nobody searches for anything. Kairos sees the work order the moment it is created, and starts gathering what the person doing this job needs to know.
- 2What it finds
Four failures, a superseded part, an ignored warning
P-101 has failed this way four times in eight years. The vendor changed the seal specification 18 months ago. Six months ago, a technician recorded a voice note about unusual vibration that nobody ever investigated.
- 3The check
Each fact is ranked by who said it
A regulation outranks a vendor manual, and a vendor manual outranks a local note. Nobody has verified the vibration note, so Kairos labels it as unverified instead of mixing it in with established fact.
- 4The delivery
A brief reaches the technician before they walk out
One screen on a phone shows the history, the part number that changed, the unverified warning, and a link to every source document behind it. They read it before they touch the pump, not after.
- 5The refusal
When the evidence is thin, it says so
Ask a safety-critical question, such as a pressure limit or a torque value. If no authoritative source supports an answer, Kairos refuses and hands you the documents instead. A confident guess at that moment is more dangerous than no answer at all.
That last step is the one people miss. A system that always answers is easy to build and impossible to trust in a plant. Knowing when to stay quiet is the harder half, and it is why everything below exists.
Answers paired with
their evidence.
Kairos already knows the plant's documents, assets and history. Every screen works the same way, so nothing is claimed without a way to check it.
Six briefs an operator per hour, because an ignored brief is worse than no brief
Governor · this hour
Dispatched 07:41
P-101 isolation changed
Rank 1 · to the duty operator
Permit to work
Bypasses the governor
Safety work is never rate-limited
Briefs
Told before you ask, but never spammed
Kairos sends you what you need to know before a job starts, and caps it at six an hour. Flood someone with notifications and they stop reading all of them. That is why plant alarm systems have a cap by standard (EEMUA-191). Permit-to-work briefs skip the cap, because safety work does not wait for a quota.
What sits under those eight screens
The eight above are screens. These are the machinery underneath them, and the cross-cutting work that has no screen of its own.
- Retrieval across mixed document types
- Elasticsearch, Qdrant and the graph queried in parallel, fused by rank, then ordered by authority
- Knowledge graph and ontology
- Six node types, six properties on every link, decades of tag aliases matched up
- Non-conformance register
- Conflicts and quarantined extractions tracked in one place, with an owner against each
- Management of Change
- A safety parameter never updates itself. Kairos drafts the change request and waits for a named approval.
- Durable ingestion
- A Temporal workflow per document, so a crash resumes at the activity that failed rather than at the first page
- Blast radius
- Change one document and see every procedure and record downstream of it that just went out of date.
- PII redaction at export
- Names are removed on the way out, never on the way in, to meet the DPDP Act 2023. Equipment tags are never matched, on purpose.
- It stops itself when it drifts
- When corrections on one class of equipment spike, nothing from that class reaches the graph until the rate settles.
Thirteen layers,
one path through.
A document goes in at one end and comes out as an answer someone can act on. Open a block to see what it is made of, and the rule it has to follow.
Built for the way plants actually run.
The plant on one screen
Live service state, open conflicts and the size of the quarantine queue, with the decisions that need a person ranked first.
No empty promises.
Provenance at 100%.
Thirty-seven questions written by domain experts across fifteen categories, run against the live stack. Higher is better.
Graded outcomes · 37 questions
- 37/37
100%
graded by fixed rules
- 37/37
100%
every answer cited
- 36/37
97%
95% confidence: 86 to 100
Retrieval
The fact reaches the model
Provenance
Sources cited
Answer quality
One valid run
One run, and every answer came from the same Llama 3.1 70B — none from a fallback model. All four misses are the authority gate refusing to answer rather than guessing, and we still grade a refusal as wrong. That understates the system, because refusing was the correct call in each of the four.
The rest of the harness
Adversarial safety
0 unsafe
15 questions written to make it guess
- Unsafe answers
- 0
- Refusals
- 12
- Answerable ones answered
- yes
- Prompt injection held
- yes
Knowledge graph linkage
18 / 21
Active vault documents with a verified knowledge edge
- Documents linked
- 85%
- Dangling provenance
- 0
- Alias resolution
- 2/2
- MDM identity
- 2/2
Proactive briefs · Layer 8
6 / 6
All graded checks pass. Soft targets are reported, not scored
- Work order raised
- 2/2
- Permit to work
- 2/2
- Tag-out · inspection
- 1/1 · 1/1
- Soft targets unmet
- 7, disclosed
Compliance gap detection
F1 0.912
Zero false positives across 52 pairs
- Precision
- 1.000
- Recall
- 0.838
- True / false positives
- 31 / 0
- Full status agreement
- 46 / 52
Entity extraction · Layer 0
F1 0.805
40 labels, 15 extractions, no fallback model used
- Precision
- 0.786
- Recall
- 0.825
- ASSET_TAG F1
- 0.889 (n=30)
- PERSON F1
- 1.000 (n=7)
Per-layer smoke
13 / 13
Every layer answered on the live stack
- Slowest read
- search 3.7 s
- Assets
- 1.7 s
- Graph / Vault
- 212 / 163 ms
- Failures
- 0
Time to a trusted answer
−9.5%
Modelled, compared against BM25 keyword search
- Traditional
- 100.0 min / 37 q
- Kairos
- 90.5 min / 37 q
- Docs opened
- 1.00 vs 1.35
- Raw machine time
- loses: 26.7 s vs 35 ms
Concurrency sweep
0% errors
2275 requests across 9 read endpoints, up to 50 users
- p50 · 1 → 50 VU
- 136 → 500 ms
- p95 at 50 VU
- 1840 ms
- Throughput
- 5.8 → 74.5 rps
- Knee of the curve
- 50 users
Soak · 60 minutes
No leak signal
Steady load against cloud stores, reads only
- Memory slope
- +8.6 MB/h
- Connections
- +4.2/h
- Errors
- 0.11% of 37,842
- Idle recovery
- 4 / 4 endpoints
Synthesis latency
p50 32.1 s
NIM 70B at the 60 second cap, quoted with its slowest cases
- p95
- 66.0 s
- Mean
- 34.1 s
- Graded questions
- 37
- Answered by
- nim 25 · openrouter 8
How every one of these is graded
Fixed rules, never another model.
An answer has to contain the facts the question requires, avoid the ones it rules out, and cite a source. Miss any of the three and it counts as wrong. The same rules run every time, so two runs of the same question are always scored the same way.
No claim without its source.
Every fact in the graph carries six properties on the link itself. That is what lets you check an answer, undo a replacement, and tell an uncertain guess apart from a verified fact.
One fact, and everything it carries
The EQ-101 seal specification was superseded on 15 January 2025.
valid_from- 2025-01-15
- When this fact started being true.
valid_to- 9999-12-31
- Still open. Closed when something replaces it, never deleted.
authority_level- 3 · OEM manual
- A regulation outranks a vendor manual, which outranks a local note.
document_id- FP-SB-2025-04
- The exact source it came from.
confidence- 0.91
- Anything below 0.7 goes to quarantine instead.
verification_status- Verified
- A named engineer signed this one off.
An illustrative fact. Every real one in the graph carries the same six.
Your questions,
answered.
The things people ask before they trust a system with a safety case. Still curious? Read the source, or open the workspace and try it yourself.
Read the sourceWhat does Kairos actually do?
It turns the documents, drawings, asset history and field notes a plant already owns into answers people can act on. Every answer shows the source it came from, how authoritative that source is, and how confident the system is.
How is this different from a chatbot over our documents?
A chatbot gives you text. Kairos gives you text plus proof of where it came from, and refuses to answer when a safety-critical question has weak evidence behind it. We grade retrieval and citations with fixed rules, not with another model.
What does it need to get started?
The documents you already have. Drawings, procedures, vendor manuals and maintenance history all go through the ingestion pipeline. Nothing has to be rewritten first.
Built with Next.js, React, TypeScript, Tailwind CSS, Python, FastAPI, Go, Celery, Neo4j, Elasticsearch, Redis, Supabase, Docker, NVIDIA, OpenTelemetry, Grafana, GitHub.