Safety document automation and phone sign-off for manufacturing sites
SAFETALK is a web service for small manufacturing sites in Korea. The person who handles safety on top of their main job gets each day’s safety documents as a draft, checks only what changed, and approves with one button. Workers read task-specific training on their own phones and sign there — no app, no login.
Pre-launch · in development
0Model calls on the daily fill path
3Checks around every model call
Row-level securityTenant isolation
Hash-chained · append-onlyApproval records
The problem
The same forms, rewritten every morning
Toolbox-meeting logs, training logs, work permits and checklists repeat daily. Usually only the date and headcount change, yet they are filled in by hand.
Gathering everyone for one signature sheet
Workers are spread across the site. Collecting signatures means pulling them into one place before work starts.
Every client has its own form
Prime contractors require their own Word, Excel or HWP forms, so the same content is copied into different layouts.
What the product does
Start from yesterday
Copy the last approved document, change only what differs. The approval screen shows changed fields first as “before → now”.
One-tap approval with a record
A single button approves. The account, device, the changes that were shown and the time are recorded automatically. Approved versions cannot be edited; corrections create a new version with a reason.
Training and signature on the worker’s phone
A per-worker link opens four to six task-specific cards, a short quiz and a signature pad. The signature sheet and training log fill themselves.
Your own forms, filled by code
Upload a Word or Excel form, AI proposes which cells are fixed, daily, auto-filled or signature fields, and a person confirms once. After that, deterministic code fills the form — no model call on the daily path.
The product itself is Korean-only. This page is an overview for partners and reviewers. Korean site
Platform
Platform architecture
SAFETALK helps the part-time safety officer at a small manufacturing site get daily safety documents and task-specific training as drafts and approve them with one button. Today the foundation is built and tested — tenant isolation, sign-in rules and the AI safety gate — and the screens and document features are being built in order.
Layers
Screens, business rules, AI, inference and data are separate layers with a single path between them. Business modules cannot see each other’s internals; automated checks enforce the boundary.
ScreensNext.js · React
Safety officer console
Worker phone view
Operator console
Business modulesTypeScript · rules as pure functions
Documents
Approval
Forms
Training
Chemicals
Document map
Output
Audit
Identity
Tenancy
AI gatewayThe single door for every model call
Personal data block
Schema validation
Number check
Evidence stored
InferenceSwappable provider slot
NVIDIA NIM
NeMo Guardrails
NeMo Retriever
Planned evaluation
DataPostgreSQL
Row-level security
Append-only records
Hash chain
File storage
How one document is made
AI is involved only at registration and drafting. It is checked on the way out and on the way in, and after a person confirms, only deterministic code runs.
Document processing recordTBM
No.
Step
What happens
01
Customer materialInput
Company forms (Word, Excel), risk assessment tables, supplier safety data sheets
02
Personal data blockOutbound check
Phone numbers are redacted; a request containing a name from the worker roster is not sent
03
ProposalAI
Form field mapping, training card drafts, safety data sheet extraction — structured output only
04
Schema and number checkInbound check
Output that breaks the schema or contains a threshold not in the verified rule table is discarded and retried
05
ConfirmationPerson
The safety officer opens the evidence, edits and confirms
06
Daily fillCode
Documents are filled from what was confirmed. No model call on this path
07
Approval and recordPerson
One button. An append-only record and a PDF hash remain
If output never passes the checks, nothing wrong is shown; the screen falls back to “please write it yourself”.
AI gateway
Whatever model is used, a call passes through one function. It is checked once on the way out and twice on the way in; output that fails never reaches a screen.
Personal data block
Phone number patterns are redacted and text is matched against the company roster. If anything remains, the request is not sent.
Schema validation
Model output is accepted only in a predefined schema. Instructions hidden inside customer documents are treated as data.
Number check
Every number with a unit is extracted and matched against the rule table. Values not verified by a person with a legal basis invalidate the output.
AI call inspection recordExample
Step
What happens
Result
request
safety data sheet extraction · prompt version recorded
Proceed
outbound
1 phone number redacted · 0 roster names
Pass
inference
structured JSON output only
Proceed
schema
valid
Pass
numbers
1 threshold not in rule table → discard, retry
Discard
attempt 2
schema ok · numbers ok · evidence linked
Pass
result
saved as draft → waits for a person to confirm
Proceed
If two attempts fail, no draft is produced and the screen falls back to manual entry.
DRAFT
Inference layer
The gateway is built so the model provider can be swapped. For a setup where customer documents never leave our infrastructure, we take a self-operated GPU server with NVIDIA’s inference stack as the reference. Three slots in the gateway line up with three parts of that stack.
AI gateway
Personal data block
Schema validation
Number check
NVIDIA NIMModel serving
Serves the extraction model as an inference service. In the gateway only the endpoint changes; the checks stay the same.
NeMo GuardrailsRules on requests and answers
A further layer on top of the three code checks, applied to the exchange with the model — for cases such as hidden instructions.
NeMo RetrieverEvidence ranking
Reranks candidate risk assessment rows and clauses when choosing the evidence a draft is grounded on.
Evaluation plan
None of the three parts is adopted yet; SAFETALK does not run on NVIDIA technology today. For each, we have fixed which product function it applies to and what we will compare against the current approach.
Evaluation plan
Product function
Form field mapping · safety data sheet extraction
NVIDIA NIMWe plan to evaluate
Where it applies
Serve an open model on a GPU server we operate and give it one document extraction task. Nemotron models are included in the comparison.
Current approach
No model connected yet. A single commercial API is the default candidate.
What we will compare
On the same synthetic documents: per-field accuracy, response time, cost per document
Product function
Safeguards around model calls
NeMo GuardrailsWe plan to evaluate
Where it applies
Apply business rules as an additional layer on requests to and answers from the model.
Current approach
Three code checks (personal data block, schema validation, number check) with unit tests.
What we will compare
Share of filtered cases: requests containing personal data, invented thresholds, instructions hidden inside documents
Product function
Evidence lookup for training card drafts
NeMo RetrieverWe plan to evaluate
Where it applies
Rerank risk assessment rows and regulation clauses to pick the evidence a draft is grounded on.
Current approach
Design stage. Each sentence is required to link to its evidence.
What we will compare
For the same task, the share of relevant evidence appearing in the top results
Why we need a GPU server
Customer documents stay on our server
Forms and risk assessments describe how a company works. Running the model on infrastructure we operate keeps those documents from going to a third-party API. First we need to know whether quality holds.
Cost per document
Our customers are small sites, so the cost of processing one document matters. We will process the same documents both ways and compare cost and response time side by side.
The workload is modest
AI runs at form registration and drafting, never on the daily fill path. We need enough compute to compare extraction quality and cost, not large-scale capacity.
How we will measure
An evaluation set of synthetic forms, safety data sheets and risk assessment tables. No real customer documents or worker data.
Field-level comparison with reference answers. Producing a value that is not in the document counts as a failure.
Cases with instructions hidden inside documents. Following them is a failure.
Until an evaluation is finished, we do not describe a technology as in use.
Record integrity
A customer must be able to show later that a document was reviewed and approved. The database itself enforces this, not application code.
Approved versions never change
A database trigger rejects updates and deletes on approved versions. Corrections need a new version and a reason.
Append-only records
Event log, approvals and signatures reject update, delete and truncate.
A hash chain
Each event log row carries the hash of the previous row. Changing one row breaks everything after it.
Tokens without originals
Session and worker-link tokens are stored only as hashes. Passwords are stored with scrypt.
Event logExample
No.
Event
Previous hash
Row hash
1
Document created
00000000
3fa91c2e
2
Sent for approval
3fa91c2e
b70d44a9
3
Approved
b70d44a9
e15c08f3
4
7 signatures stored
e15c08f3
9a26d7b0
Each row includes the hash of the previous row. Changing one row invalidates every hash after it.
Implementation status
We keep three lists so that what exists is never mixed with what is planned.
Built and tested
Code exists and automated tests pass
Tenant isolation
PostgreSQL row-level security keeps rows outside a company’s scope invisible. Verified by integration tests against a real database.
A single AI gate with three safeguards
Every model call goes through one function: personal data block, schema validation, and rejection of numbers not in the rule table. Implemented in code with unit tests. No model is connected yet.
Sign-in rules
Password verification (scrypt) and session open, check and close.
Business rules and structure tests
Approval, versioning and permit rules are pure functions; tests enforce module boundaries and the approved dependency list.
In progress
Designed; being implemented in order
Sign-in screen and API
Real accounts are created only after rate limiting is in place.
Document engine and approval inbox
Copy from yesterday, changed fields, one-tap approval, bundled PDF.
Worker training and signing
Link delivery, phone cards and quiz, signature, on-site mode.
Company forms and chemicals
Field confirmation and filling; registers, handling guides and warning labels from safety data sheets.
To evaluate
Not started. Only the target and the metric are fixed
Serving models on our own server
Run a document extraction task on NVIDIA NIM and measure quality, latency and cost.
A second layer of safeguards
Compare filtering rates with NeMo Guardrails added to the existing code checks.
Reranking evidence
Apply NeMo Retriever to evidence lookup for training card drafts.
NVIDIA, NIM, NeMo and Nemotron are trademarks of NVIDIA Corporation. They are named here to identify what we plan to evaluate and do not imply partnership or endorsement.