Manual visual checks are tiring and inconsistent
Operators inspect hundreds of near-identical parts per shift. Attention drops, and a missing clip looks a lot like a present one at the end of a long day.
Pre-MVP · founder-led concept
QualiEdge AI Visual assembly inspection.
An AI-assisted visual assembly inspection concept for small and medium manufacturers. Operators stay in control: every part gets a PASS, FAIL or REVIEW decision — with the evidence behind it.
- synthetic scenarios
- 5
- synthetic scenarios
- decision states
- 3
- decision states
- cloud uploads required
- 0
- cloud uploads required
- only, offline-first
- CPU
- only, offline-first
DECISION
PASS — 4/4 visible checks · SYNTHETIC
PRE-MVP · SYNTHETIC DATA · NO FUNDING CLAIMED · OFFLINE-FIRST · HUMAN-IN-THE-LOOP
01 · The challenge
Small manufacturers still rely on tired eyes.
Frequent changeovers make rigid vision costly
Small-batch producers switch variants often. Classic vision systems need specialist reconfiguration for every new product version.
Rare defects mean few training examples
Good factories produce few bad parts. That makes purely data-hungry deep learning hard to train and harder to trust.
Problem hypotheses to be validated through customer interviews.
02 · The solution
A transparent workflow, not a black box.
Every decision is traceable to a region of the image and a versioned rule. When the system is unsure, it says so.
- 01
Capture
Fixed camera, controlled light, image quality gate.
- 02
Analyse
Versioned recipe of visible feature checks in fixed ROIs.
- 03
Review
Uncertain or poor captures go to an operator — never guessed.
- 04
Record
Evidence image, timestamp and recipe version, append-only.
V1 scope
- Fixed camera + controlled light
- Image quality & alignment check
- Versioned recipe of visible feature checks
- PASS / FAIL / REVIEW decisions
- Evidence image + timestamp
- Auditable human review (append-only)
V1 excludes
- PLC writes
- Automatic machine rejects
- Safety-critical parts
- Cloud upload of customer images without approval
03 · Simulator
Break the part. Watch the checks respond.
A deterministic, in-browser simulator of the recipe-based baseline. The detector only reads pixels — it never sees the scenario label.
04 · Proof of concept
A reproducible baseline, honestly labelled.
Python + OpenCV on synthetic images. No model has been trained — this is a starting point for measurement, not a result.
05 · Business model (hypotheses)
Simple, per-station pricing — to be tested.
Feasibility study
€2,000
one-off
Sample review, lighting & fixture trial, written go/no-go.
Installation & integration
€4,000–€7,000
per station
Camera, light, fixture, recipe setup and operator training.
Software subscription
~€1,200
per station / month
Recipes, updates, evidence store, review dashboard, support.
Pricing hypotheses, not quotes or signed commitments.
Illustrative ROI calculator
- Annual benefit
- €39,840
- First-year cost
- €19,400
- Net (year 1)
- +€20,440
benefit = defects × cost + hours × rate × 240 · cost = €1,200 × 12 + €5,000. Illustrative only; assumes 240 working days and €5,000 installation.
06 · Validation
What we will measure — before claiming anything.
Defect recall
How many defective parts are detected.
Result: not yet measured
False reject rate
Good parts wrongly rejected.
Result: not yet measured
Cycle-time latency
Does inspection keep up with line speed?
Result: not yet measured
Changeover effort
Hours to configure a new product variant.
Result: not yet measured
These are metrics we intend to measure on permissioned real samples — no results are claimed yet.
07 · Customer discovery targets
A funnel of goals, not achievements.
Focus: quality managers and production leads at small and medium manufacturers in Baden-Württemberg.
30relevant manufacturers
10interviews
3pilot discussions
1paid feasibility study
Goals, not achievements
08 · Roadmap
Four gates. No skipping.
Each gate depends on the previous one.
- 1
Gate 1 · current
Verify university eligibility & market need
- 2
Gate 2
Table-top prototype & evaluation
- 3
Gate 3
Real permissioned component images & failure tracking
- 4
Gate 4
Limited factory pilot — only if technically feasible and scoped
09 · Collaborate
Help shape a better inspection workflow.
- A German university host to assess EXIST Startup Grant eligibility
- Computer-vision mentorship
- IP and evaluation advice
- Introductions to manufacturers with visual inspection pain
There is no EXIST approval or university sponsorship at present.
Ammad Aslam