From site visit to approved client report, with every statement traceable.
Inspectors capture photos, voice notes and typed observations on their phones. Slomane turns them into a structured draft. A named reviewer checks each point against its evidence, approves a version, and downloads the report.
Built for small engineering consultancies and supervision teams. First report type: routine building site visit report.
Block B · Draft for reviewExample data
Photo
Cable containment incomplete, east corridor.
“Containment installation is incomplete in this section.”
Supported by source
Photo
Pipe sleeves reported as 16 mm by the foreman.
Reported, not measured by the inspector.
Supported by source
Photo
Possible honeycombing, suspected.
Which column? Location not stated.
Needs clarification
Draft, not approved1 waiting for a reviewer
Never guessesmissing stays missing
Never certifiesno compliance judgments
Always shows evidencephoto and source beside each point
Named approvaltied to an exact version
PRODUCT WALKTHROUGH
Watch a site note become an approved report.
A two-minute screen recording of the working demo, from capture to approved Word file. It uses an example site note, and the AI step is the labelled replay of a recorded model response. Press play or pick a chapter.
HOW THE AI WORKS
Nine steps from a messy note to a locked report.
The model does one job: structure what was said. Code checks its work, and a person approves. Step through the pipeline to see what goes in, what the AI does, what guards the step and what comes out. Tags show what runs in the live demo and what belongs to the pilot build.
FROM THE FIELD
Evidence stays attached to what the camera saw.
Each photo, note and location is tied to the observation it belongs to, so the report can show where every statement came from. Illustration only.
AI-NATIVE
Built around the model, not bolted onto a form.
A typical “AI feature” is a button that rewrites text inside a document template. Slomane starts from the other end. The model’s job, the checks on what it says and the person who signs off are the product. This is what that means here, and what already works.
1
Structured by contract
The model returns findings in 11 fixed fields and follows nine rules. Anything the note does not state comes back empty.
2
Evidence first
Every finding carries a quote from the inspector’s note. No quote, no finding.
3
Checks outside the model
Plain code, not the AI, verifies each quote, flags missing locations and marks reported measurements.
4
A person decides
A named reviewer approves one exact, locked version. The AI never approves its own work.
What is different in practice
Question
A typical bolt-on AI button
Slomane
Where the AI sits
Rewrites text inside a template.
First step of the pipeline: a raw note becomes structured findings.
What it may say
Whatever reads well.
Only what the note states. Missing details stay empty.
How its output is checked
The author reads it again.
Code checks every source quote against the note.
When it is wrong
Easy to miss.
The finding shows “Quote not found” and cannot be approved until it is excluded with a reason.
Who is accountable
Unclear.
A named approval on an exact version, with a SHA-256 fingerprint.
Changing the model
Rebuild the workflow.
The model call sits behind one function. The schema, checks and review stay the same.
Working in the MVP
Tested live
Structured extraction, source-quote checks, missing-location flags, the review gate, named approval, and Word and JSON export. See the test ›
Still to prove with real material
Pilot
Voice transcription, photo matching and accuracy on real Nigerian English and Pidgin recordings. These are tested in the pilot, not assumed.
Pilot modules are in the first build. Planned and Later are not promised for the pilot.
USE IT · YOUR SITE VISIT
Use it on your own site visit.
Add a note and photos from a real visit and press Create draft. Claude structures the note, automated checks verify every source quote, you attach photos to each finding as evidence, resolve what was flagged, approve under your name, and download a Word report with the photos in it. Your note and photos stay in this browser: they are not sent to a server, and photos are never sent to the AI.
Add your note and photos, then Create draft
Review: resolve flags, attach photos
Approve with your name
Download the Word report
Up to 6,000 characters for each draft. Write each observation on its own line if you can.
Photos from the visit0 of 20
or drop images here. JPG or PNG, up to 20. Each photo is shrunk in your browser and its location data is removed. Photos are not sent to the AI or to any server.
Saved in this browser so you can come back.
Run log
Findings
–
Quotes verified
–
Flagged for a human
–
AI extraction time
–
Characters in → out
–
Machine time, note to draft
–
Guardrail tests
Runs the real verification code from this page on fixed test outputs, including bad ones. It checks the safety rules, not model accuracy.
Working in this demo
AI structures a free-text note into findings. Without a live AI connection, your note is split into findings as written.
Photos you add are shrunk in your browser, attached to findings by the reviewer, and placed in the Word report.
Every source quote is checked against the note automatically.
A missing location blocks approval until a person fills it in.
A named approval freezes the version, with a content fingerprint.
The Word report and JSON snapshot come from that frozen version.
Still to prove at scale
Voice transcription, and photo upload tested on real phones.
Saving visits and photos on a server for each firm, across devices.
Many firms at once, each fully isolated, under load.
Accuracy on real Nigerian English and Pidgin recordings.
The MVP is the working demo on this site: one report type, a routine building site visit report, and one full path from a site note to an approved Word file. The pilot build adds phone capture, photos, voice and a firm’s own template. Nothing in the last column is promised for the pilot.
Working now · MVP
The full path, end to end
AI structures a note into findings. Without a live AI connection, the note is split into findings as written.
Photo upload from a phone or computer. The reviewer attaches photos to findings, and they go into the Word report. Photos stay in your browser.
Every source quote is checked against the note.
A missing location blocks approval.
An unsupported finding must be excluded with a reason.
Named approval freezes the version with a fingerprint.
Word report and JSON snapshot come from that version.
Pilot build
What firms get first
Company-isolated projects and visits.
Capture on phones with the camera, and a location the inspector chooses.
Visits and photos saved on a server for each firm, across devices.
Voice transcription, once options are compared on real recordings.
Word report in the firm’s standard template, setup included.
Role-based approval and an audit trail.
Later · not promised
After the pilot
PDF export, once the Word flow is proven.
Offline capture, if connectivity blocks real visits.
Other report types.
Pins on a real building model. The 3D view today is an illustration.
How the MVP will be judged in the pilot
These are proposed measures to agree with each firm. They are not results.
Median time from visit to a correct report, against a baseline measured before the pilot. The proposed target is 40% less hands-on time.
How often reviewers change or exclude a finding.
How often the automatic checks flag something a person then confirms was wrong.
Whether reviewers are willing to approve from the screen, with evidence beside each finding.
Site evidence is scattered. The report shouldn’t be.
After every visit, photos sit in a gallery, notes in chat and measurements in memory. Someone spends the evening matching them up. Slomane keeps each observation together and turns the lot into one draft that a person approves.
What comes back from a visit
Photos14
Voice notes3, transcribed
Typed notes2, as written
Measurements6, original wording
Locations3, picked by the inspector
One visit record
What the team gets
Structured draftgaps flagged
Evidence linksphoto and quote per finding
Review queue1 flagged for a person
Approved reportWord from the locked version
Counts come from the example visit in the demo.
LIVE 3D
Watch a building become a reviewed record.
A real-time 3D scene drawn in your browser. The building rises as a visit is captured, the AI scans it and places findings, a reviewer resolves them, and an approval locks the version. Drag to turn it. Example building.
Paste a site note. Watch it become structured findings.
This calls Claude with the rules from the build plan: nothing is guessed, missing values stay empty, and every finding needs a quote from your note. Your note is sent to Claude through this site’s server for this request only. Nothing is stored.
The rules the AI must follow
Use only what the note says. Anything not stated is null. Keep measurements in the original wording and say who measured. Keep suspected issues worded as suspected. Never turn “not inspected” into “no defects”. No diagnosis, certification or code judgments. Every finding needs a quote copied from the note, and the page checks that the quote really exists.
RAW TO STRUCTURED
Drag to see what the AI does to a site note.
Left of the handle is the structured draft, right is the raw note. Use the arrow keys if you prefer. Example data.
RAW NOTE
Block B first floor east corridor, containment still not finished, electrical contractor to complete before ceiling closure. Plant room, foreman says the sleeves are sixteen millimetres, I haven’t measured. Looks like honeycombing on one of the columns, need to check it.
STRUCTURED DRAFT
Containment incompleteSupported
Block B, first floor, east corridor
Sleeves reported 16 mmSupported
Reported, not measured
Possible honeycombingClarify
Location not stated
VISIT COPILOT
Ask your visit.
Copilot answers only from this visit’s findings, cites them, and says so when they can’t answer. It reads the same findings as the review screen, including any you add.
Findings
0
Needing attention
0
Resolved
0
Needs attention
Ask Copilot
HOW THE AI STRUCTURES A VISIT
From a messy note to an approved report, step by step.
A looping animation of one example observation moving through the product: the AI reads the note, fills fixed fields, checks itself, a reviewer resolves what it flagged, and the approved version is locked.
1 · Inspector’s note (voice or typed)
Block B, first floor, east corridor: containment still not finished, electrical contractor to complete before ceiling closure. Plant room, foreman says the sleeves are sixteen millimetres, I haven’t measured. Looks like honeycombing on one of the columns, need to check it.
FindingCable containment incomplete
LocationBlock B, first floor, east corridor
ActionContractor to complete before ceiling closure
✓ Quote found in note
FindingSleeves reported as 16 mm
Measured by inspectorNo, reported by foreman
Actionnull
✓ Quote found in note
FindingPossible honeycombing, suspected
LocationnullBlock B, column C4
⚠ Location not stated. Sent to reviewer
✓ Location entered by reviewer
✓ Approved R1
Site Visit Report · Block B
3 findings · 1 reported measurement marked “not measured” · 1 suspected issue worded as suspected · approved by a named reviewer · Word generated from this exact version
Illustrative animation with example data.
THE SIX STAGES
Watch a site visit become a reviewed report.
An animated preview of the six stages, built from the product plan. It is not a screen recording of finished software. Pick a stage or let it play.
One workflow, four steps
The inspector supplies the engineering judgment. The software preserves and organizes it.
01
Capture
One observation at a time: location, photos, a short voice or typed note, and any stated action.
02
Draft
Voice is transcribed and facts are extracted into a strict schema. Missing values stay empty, never guessed.
03
Review
Findings sit beside their photos and source note. Unclear points are flagged for the reviewer.
04
Approve and export
An explicit, recorded approval freezes a version. Word and PDF come from that same version.
THE EVIDENCE CHAIN
Every statement links back to where it came from.
PhotoIMG-0041
→
LocationBlock B · Level 1
→
Observation#3 in the visit
→
Source quoteexact words, checked
→
Findingsuspected, flagged
→
Approved versionR1 · named reviewer
BUILT TO SCALE · PLAN
It works today. Here is what it takes to run it for many firms.
Each site visit is one independent background job, and every company’s data is kept apart. Adding firms means running more of the same job, not building a different product. What still has to be proven is volume, reliability and accuracy on real material. That is the work extra cloud capacity would pay for.
01
One visit, one job
Processing runs in the background. A retry rebuilds the draft from scratch, so a failed run never duplicates or loses work.
02
One bad note can’t sink a visit
If one observation fails, the rest still process and the reviewer gets a clear placeholder to fill in by hand.
03
Isolated by design
A company id on every table with row-level security, and approved versions that cannot be edited afterwards.
04
Replaceable AI provider
The model call sits behind one function, so moving between providers does not touch the schema, the checks or the review flow.
Planned architecture on Azure
Planned, not deployed The pilot plan uses Next.js, Supabase, Trigger.dev and Vercel. This is the proposed mapping for scale. The AI step in this demo runs on Claude.
Container Apps jobs or Functions, fed by a Service Bus queue
Speech to text
Provider chosen by testing Nigerian English and Pidgin
Azure AI Speech, tested on the same recordings
Structuring
Strict-schema model call, replaceable
Azure OpenAI with the same schema and checks
Database and sign-in
Supabase Postgres with row-level security
Azure Database for PostgreSQL and Microsoft Entra ID, policies adapted
Monitoring
Job ids, durations, model versions, cost metadata
Azure Monitor and Application Insights
Size it
Move the sliders. This is arithmetic from your inputs, not a benchmark or a quote.
Reports a month
–
AI calls a month
–
Busiest-hour requests per minute
–
Parallel workers at peak
–
Photos stored a month
–
Audio a month
–
AI tokens a month (rough)
–
Assumes 22 working days, one AI call per observation, about four characters per token, and photos only for storage. Replace every assumption with measured values from the pilot logs.
What extra capacity would fund
A
Evaluation on real material
Run 30 complete visit packets, 10 of them held out, through transcription and extraction, including Nigerian English and Pidgin recordings. Target: at least 95% precision and recall on labelled facts. A target, not a result.
B
Load and isolation tests
Deliberate cross-company access attempts, double clicks on Generate, provider timeouts and failed uploads. Targets: every attempt rejected, and at least 95% of jobs finish with no developer help.
C
Recovery, actually rehearsed
Restore both the database records and the real photo and audio files after a drill, within a recovery target agreed in advance.
D
A controlled pilot
Two or three firms, about ten real reports each, measuring time saved, correction effort, job reliability and processing cost.
Slomane isn’t a chatbot that writes a report. It structures what was observed and checks its own work before a person sees it.
01
Transcribe
Voice notes become text, kept with the original audio.
02
Extract
Findings must fit a strict schema. Anything that doesn’t is rejected and retried.
03
Verify
Each finding’s source quote must exist in the input, or it goes to the reviewer.
04
Flag
Missing locations, vague wording and suspected issues are marked, never resolved by guessing.
05
Freeze
Approval locks an exact version. Exports come only from that snapshot.
WHO IT’S FOR
Built for the people who sign off on site reports.
Inspectors
“Can I capture this on site and not write it up tonight?” One observation at a time, from a phone.
Reviewers
“What did they actually see?” Photos and source notes sit next to every finding.
Firm principals
“Who approved this, and when?” A named approval tied to an exact version.
Clients
“What was and wasn’t inspected?” Reports say so plainly instead of implying all clear.
First report type: routine building site visit report. Other types come after the pilot.
A DAY WITH SLOMANE
See it from each person’s seat.
THE SHIFT
What changes for the team
Area
Today
With Slomane
Matching evidence
Photos in a gallery, notes in chat, matched by hand in the evening.
Each observation keeps its photos, note and location together.
Measurements
Remembered, or reworded on the way into the report.
Original wording kept. A figure someone else reported is marked as not measured by the inspector.
Missing details
Filled in from memory so the report reads well.
Left empty and flagged for the reviewer.
Review
The reviewer can’t easily tell what was seen from what was assumed.
Every finding sits beside its source quote and photos.
Approval
Emailed Word files, and no clear record of which version was signed off.
A named approval on an exact, locked version.
Assembling the report
Copy and paste into the firm’s template.
A Word file generated from the approved version.
The time-saving target is a hypothesis to measure in the pilot, not a result.
REAL CASES
Why site records matter: what public findings say about real collapses.
These are public cases. Slomane had no part in any of them, and we do not claim it would have prevented one. We cite them because the findings keep returning to the same few questions: what was approved, what was built, and who was checking. The site visit record is where those answers should be written down.
Public sources, linked belowNot Slomane customers
LAGOS · 1 NOVEMBER 2021
The Gerard Road collapse, Ikoyi
A 21-storey building under construction at Gerard Road, Ikoyi, collapsed. These findings were reported publicly.
The coroner’s inquest found design errors and a combined stress failure of the columns on the lower floors, no quality assurance or quality control, and poor project management on site.
It found the approving and supervising agencies negligent. The project board did not display the names and addresses of the professionals involved.
The project’s structural engineer told a panel that the brief was changed more than once during design. He said he never saw an approval, and that his team worked on site from copies of its own drawings.
A pile integrity test found failures in 23 piles, and he recommended replacing them.
The records question: which drawing was being built, who checked each stage, and who put their name to it.
NIGERIA · 2024
47 recorded collapses in one year
The Building Collapse Prevention Guild recorded 47 building collapses across 14 states in 2024, as AllAfrica reported in January 2025. Lagos had 13 and Abuja had 6.
The report, as quoted, did not give a cause for each collapse.
A real estate expert quoted alongside it named unqualified professionals, poor construction practices and substandard materials as common factors in collapses in general. He said regular checks could help find cracks and leaks early.
The records question: a count shows how common the problem is, not what went wrong in each case. Only a record made at the time can show that.
What Slomane does and does not do about this
It does
Keep each observation beside the inspector’s own words.
Mark a figure that was reported by someone else and not measured by the inspector.
Flag a missing location and hold approval until a person supplies it.
Lock one version under a named approval, with a fingerprint of its contents.
It does not
Check a building against its approved drawings.
Test materials, piles or concrete.
Judge whether a defect is acceptable.
Certify structural adequacy, or replace the engineer on site.
Real visits: none published yet
Slomane has not yet been used on a real client’s visit. The cases above are public records, not our results, and the live demo runs on an example note. When a firm completes a pilot, we plan to publish what was measured, including what did not work, with that firm’s permission. If you have a real site note, you can run it through the demo now.
A routine visit, from first note to signed report.
This is how a routine visit is meant to go, told step by step with the roles involved. It describes how the product is used. It is not a customer story, and it uses the same three observations as the live demo, so you can run them yourself. Tags show what the MVP does today and what belongs to the pilot build.
Same three observations as the demoNot a customer story
The site engineer walks the block
A site engineer at a contractor is on a residential project. In the first-floor east corridor she sees cable containment that is still unfinished. She takes two photos and types a short note, with the location.
Pilot build Capture one observation at a time on a phone, with a location the inspector chooses. The MVP shows this step as a preview.
A figure she did not measure
In the plant room the foreman says the pipe sleeves are sixteen millimetres. She has not measured them and writes it as she heard it.
In the MVP The wording is kept, and the figure is marked “reported, not measured by inspector”. A foreman’s number never becomes the inspector’s measurement.
A doubt, with no location
She notes some honeycombing on one of the columns, “need to check it properly”. She forgets which column.
In the MVP The draft says “possible honeycombing, suspected” and leaves the location empty and flagged. Suspected stays suspected, and a missing location blocks approval.
The draft is waiting
Back at her desk, she finds the three notes already structured, each finding beside its source quote and photos. She enters the missing column, then ticks each finding against its quote.
In the MVP Code has already checked that every quote appears in her note. Approval stays locked until each finding has a location and a source check.
If the AI had invented something
Suppose the draft had added “rebar cover inadequate”, which she never wrote. Nothing in her note backs it up.
In the MVP That finding shows “Quote not found” and cannot be ticked as checked. It has to be excluded, with a written reason. We tested this case. See the test ›
The reviewing engineer signs off
The reviewing engineer sees what was observed beside what was only reported, and approves under his own name. Version R1 is locked and a Word report is generated from it for the client.
In the MVP A named approval, a SHA-256 fingerprint, then Word and JSON export. A later correction starts a new draft that needs approving again.
What this walkthrough does not claim
It does not claim a time saving. That is measured in the pilot, against a baseline.
It does not show accuracy on real recordings or real sites. That is also for the pilot.
The product makes no engineering judgement. It does not decide whether honeycombing is acceptable or whether a sleeve size is right.
One site note, tested live from capture to approved report.
On 10 November 2026 we ran the example note through this deployed demo and wrote down what each stage actually did. This is a product test on an example note, not a customer result. The AI step was a replay of a recorded model response, so the test covers the checks, the review, the approval and the export. It does not test how accurate the AI is.
Example noteRun on the deployed siteAI step: recorded replay
The note
EXAMPLE INSPECTOR NOTE · 328 CHARACTERS · 51 WORDSFirst floor east corridor in Block B: cable containment installation is still incomplete, electrical contractor to complete before ceiling closure. In the plant room the foreman says the pipe sleeves are sixteen millimetres, I haven’t measured them. Looks like some honeycombing on one of the columns, need to check it properly.
What came back
3findings returned
3 of 3source quotes found in the note
1flagged for a person: location missing
87 msfor intake and checks, with the AI step replayed
Finding
Quote check
What the checks did
Cable containment installation is incomplete.
Verified
Kept the location and the contractor’s action as the inspector stated them.
Possible honeycombing on a column, suspected.
Verified
No location in the note, so it was flagged and left empty. The reviewer had to supply one. The wording stayed “suspected”.
Pipe sleeves reported as 16 mm.
Verified
Marked “reported, not measured by inspector”. The foreman’s figure was not turned into an inspector’s measurement.
Review, approval and export: what we tried
Step
What we did
Result
1 · Before reviewing
Looked at the Approve button straight after the run.
Locked: “3 findings still need a location and a source check.”
2 · Partial review
Ticked the source check on two findings.
Still locked, with one finding left.
3 · Missing location
Ticked the third check but left its location empty.
Still locked.
4 · Location added
Typed a location for the flagged finding.
Unlocked.
5 · No name
Pressed Approve with the reviewer name empty.
Refused, and asked for a reviewer name.
6 · Approve
Entered a reviewer name and approved.
Revision R1 frozen. Every field locked. A SHA-256 fingerprint was recorded.
7 · Edit after approval
Tried to change a location.
The field was disabled.
8 · Export
Generated the Word report and the JSON snapshot.
Word: a three-row table of about 8 KB. JSON: about 1.3 KB. The reported measurement carries measured_by_inspector: false, and the flagged finding is recorded as reviewer_confirmed.
9 · Fingerprint check
Recomputed the SHA-256 of the exported snapshot outside the demo.
Matched the stored fingerprint.
10 · Tamper check
Changed one word in the first finding (“incomplete” to “complete”) and hashed again.
The fingerprint no longer matched.
A second test: what if the AI states something the note never said?
We deleted the honeycombing sentence from the note (79 characters) and replayed the same recorded response, as if the AI had reported something the inspector never wrote. The checks caught it: 2 of 3 quotes found, and the honeycombing finding showed “Quote not found”.
The test also found a gap in our demo. In the first run, a reviewer could still tick “Checked against source quote” on that finding and approve it. We changed the demo so that a finding whose quote is not in the note can no longer be marked as checked. It has to be excluded with a written reason, and approval stays blocked if every finding is excluded. We ran the test again: the excluded finding is left out of the report, listed in the Word file and the JSON snapshot with its reason, and changing the reason changes the fingerprint.
What this test does not show
No real inspector, site or customer was involved.
It is one note and one run on an example note, not on a real visit.
The AI answer was a replay, so the live model’s accuracy was not tested.
Time saved was not measured. The time-saving target is a hypothesis for the pilot.
The Word file is a plain table, not a firm’s own template, and it makes no compliance judgement.
Each company keeps its own projects. Start a new visit from a project, and the capture step below is tagged to it. These are examples and live only in this preview.
Your projectsExample data
Try the capture step
This is the inspector’s view: one observation at a time, with a location, a short note and photos. Photos you add stay in this browser and are not uploaded to a server. Saved observations appear in the review screen below.
New observationPreview only
0 of 20 photos
Try the review step
This is a front-end preview of the reviewer’s screen, using an example visit. Nothing is sent to a server or to AI here. Edit a finding, resolve the flagged point, then approve.
Block B – Routine site visit
Ref. DEMO-0001 · Example data
Example data
3D SITE VIEW
See where every finding sits in the building.
Drag to turn the model, or use the arrow keys. Pins come from the review screen: blue to check, green resolved, amber location not stated. Example building, a schematic and not a survey.
Findings in the model
Select a pin or a finding.
Estimate your own time
Set the sliders to your firm’s numbers. The reduction is an assumption you control. 40% is the pilot’s proposed target, not a measured result.
Hours per month now
0
Hours with Slomane
0
Hours back each month
0
SECURITY
Client site data stays protected.
These are the pilot’s design requirements. Each is tested before any firm uses the product with real work.
Company isolation
Enforced in the database, so one firm can’t see another’s records.
Private files
Photos and audio sit in private storage, shared only through short-lived links.
Immutable approvals
Approved versions can’t be edited. Corrections create a new draft.
Role-based approval
Only reviewers and admins can approve a report.
Audit trail
Processing and approval events are logged.
Built around honesty about evidence
Every material statement links to a source or is marked as reviewer entered.
“Not inspected” is never turned into “no defects found”.
Measurements keep their original wording beside any parsed value.
No weather, completion percentages, grades or signatures are invented.
Approval refers to the exact content version the reviewer saw.
Later corrections create a new draft that needs approval again.
Company data is isolated, files are private, and uploads show “Uploaded” only once confirmed.
No diagnosis, certification or code-compliance judgments.
EARLY ACCESS
Start with one project.
1
Talk
A short call about how your team writes site reports today.
2
Bring a real visit
We turn one of your recent visits into a draft, then review it together.
3
Run a pilot
One firm, a few users, agreed success measures and weekly check-ins.
TEAM
The people building Slomane
A note from the co-founders
A site visit report shouldn’t depend on someone’s memory at the end of a long day. We are building Slomane so that what an inspector sees and says on site travels, with its evidence, all the way into the report a client receives.
We hold ourselves to three rules. The software organizes what was observed and never adds to it. Anything it is unsure of is flagged for a person. And a named engineer approves every report before it goes out.
We are a small team preparing a first pilot with engineering firms. If you inspect, supervise or review site work, we would like to hear what your reports cost you today.
No. It structures what the inspector said and photographed. It does not diagnose, grade severity, certify, or judge code compliance. A named reviewer approves every report.
What if the AI gets something wrong?
Each finding sits beside its photos and source note, and unclear points are flagged. The reviewer can edit anything. If a correction is needed after approval, it creates a new draft that must be approved again.
Does it work with Nigerian accents and Pidgin?
That has to be tested, not assumed. The pilot plan compares transcription options on real recordings from real sites before choosing one.
Is this a finished product?
No. It is an MVP: a working demo of the whole path from a site note to an approved Word report, for one report type. The pilot adds phone capture, photos, voice and a firm’s own template. The MVP scope section lists what is built, what is in the pilot build and what comes later.
What does “AI-native” mean here?
The model’s output is the first step of the workflow, and everything around it is built to check it: a fixed schema, a source quote for every finding, automatic checks written as plain code, and a named person who approves. It is not an AI button on a document template.
Are the cases and data on this site real?
The cases under “Real cases” are public findings, with links to their sources. They are not Slomane customers. The notes and findings in the live demo are examples written for it, and they are labelled as example data. No firm has used Slomane on a real visit yet. When one does, we plan to publish what it shows, with that firm’s permission.
Where do my photos go?
Photos you add stay in your browser. The page shrinks each one, removes its location data, and keeps it in your browser’s storage until you remove it or clear the visit. They are not sent to the AI or to a server. They go into the Word report when you approve, and clearing your browser data deletes them.
What about sites with poor connectivity?
Offline capture is deliberately deferred. The pilot will show whether connectivity blocks real visits, and if it does, offline capture becomes its own milestone.
Is our data kept separate from other firms?
That is a design requirement: company data is isolated, files are private, and isolation is tested before any pilot. This preview page stores nothing on a server. Notes and photos you add are kept only in your own browser.
Controlled pilot, Abuja
OFFER TO TEST
NGN 50,000
Four-week paid pilot per company, up to 5 users and 40 reports, including standard template setup.
LIMITS
20 photos · 10 min audio
Provisional per-visit limits. Nothing beyond them is silently dropped.
GOAL
40% less hands-on time
Proposed target for median time from visit to correct report, against a measured baseline.
The price and targets are proposals to validate with design-partner firms, not established results. Want to bring one of your own recent visits to a demo? Get in touch with Emmanuel Ofodile.
Know what was seen, not just what was written.
Try the review step on the example visit, then bring one of your own.
