Drop a chart. Get a read.
The extraction layer is what this page is about — measuring the candles and the axis before any reasoning starts. It is the part of ORIN still being built, and until it ships the honest answer is a symbol, not a screenshot.
Reasoning is only allowed to start after step 5. That order is the whole argument — and it is why the feature is not shipped yet rather than shipped guessing.
Where the vision layer actually stands
Drag, paste with ⌘V, or use your camera roll. You will get the honest answer rather than a loading animation over an invented read.
Drop a chart screenshot
Drag it here, paste with ⌘V, or tap to pick from your camera roll — and see exactly what ORIN can and cannot do with an image today.
Choose an imageProcessed in your browser · never uploaded from this page
Your image is processed in the browser and never uploaded. Nothing reads it yet — the sample plan shown alongside is a worked example on a different symbol, labelled as one.
One chart. A general model, and a measured one.
The left column is what a general model returns on an NVDA screenshot today. The right is the standard the extraction layer has to meet before it ships — every line answerable with a measurement, or it does not get written.
General-purpose AI
UngroundedReads the chart as a picture and describes it
"Price is showing strong bullish momentum with support around the $126 level."
There is no $126 level on this chart. The nearest swing low is 124.10.
"I'd look for a stop below $125 and a target near $140."
Round numbers, not structure. The target sits beyond the visible range.
"RSI appears to be around 60, suggesting room to run."
No RSI panel is present in the image.
No invalidation reasoning, and no statement of confidence.
ORIN's standard
MeasuredExtract the axis, then reason over measurements
Sell-side liquidity swept at 124.10 and reclaimed within two candles.
Every level has to trace to a coordinate on the image, or it is not quoted.
Invalidation at 124.90 — beneath the order block that has to hold.
Derived from structure, not from a round number.
Target 1 at 135.91, giving 1:2.1 against the stop.
Next structural objective; risk-to-reward stated, not implied.
Confidence 86 — with the resolved sample behind that band attached, or nothing at all.
A confidence figure without its sample is a mood. The sample is still collecting.
General AI describes charts. The bar here is measuring them — which is why this one is not shipped yet.
How ORIN will read a screenshot
Three stages, and each one has an honest edge case that has to be surfaced rather than hidden. Publishing the edge cases before the feature is the point — they are why it is taking this long.
Extract
Candle bodies and wicks are isolated from the plot area, and the price axis is read to map pixels onto prices. This is the step a general model skips.
Edge case: a cropped axis means prices can only be inferred relatively. When that happens you get an "approximate levels" warning on the read — not a confident number built on a guess.
Measure
Swing points, ranges, order blocks and gaps are computed from the extracted series — the same measurements the live scanner runs, on pixels instead of a feed.
Edge case: heavily annotated charts can hide candles behind drawings. Obscured regions are marked as low-confidence rather than interpolated.
Reason
Only now does anything get written. Each line of the thesis cites a measurement, so a claim you disagree with can be traced to the number behind it.
Edge case: an image alone has no live price. If the symbol is not detected, the read is explicitly historical — use the scanner for a current one.
When a screenshot isn't enough
An image is a frozen moment, and these are the limits that survive even a perfect extraction layer. They are written as commitments the feature has to honour on the day it ships — publishing them first is how you can hold us to them.
No live price context
Unless the symbol is detected in the image, the read describes the chart as it was captured. If the screenshot is two hours old, so is the analysis — and price does not care that you only just asked.
Cropped axes limit precision
If the price scale is cut off, levels can only be placed relative to each other. The read says so rather than inventing absolute numbers, which means the warning is a feature and not a failure.
Low-resolution images lose wicks
A compressed screenshot from a chat app can lose the wick detail that structure depends on. Sub-800px images produce a lower-confidence read, and the confidence number reflects it.
For a live symbol, use the scanner
If the ticker is one we cover, the scanner pulls current data and gives you current levels — strictly better than an image of the same chart. This page exists for everything that is not that.
The honest recommendation: use the AI chart scanner when you have a symbol, and this page when you have an image. Pointing you at the other tool costs us nothing and saves you a worse answer.
Questions traders actually ask
It depends entirely on whether the model measures the image or describes it. A general model reads a chart as a picture and produces plausible-sounding prices. Doing it properly means extracting the candles and the price axis first, so every level quoted traces back to a pixel actually found — which is why ORIN has not shipped it yet rather than shipping the describing kind.
You can, and it will answer confidently. The failure mode is that general models infer prices from context rather than reading them, so levels drift — often by amounts that matter to a stop. ORIN takes the symbol instead and reads real bars, which is a narrower job done properly rather than a broader one done approximately.
You can upload one, and ORIN will tell you it cannot read it. The vision layer is the one part of the pipeline still being built, and guessing levels off an image is the specific failure the product is designed around. Type the ticker and the same analysis runs on real candles.
They will be flagged as yours rather than folded into the read. A line you drew is evidence about what you were thinking, not evidence about the market — keeping those separate is the difference between analysis and agreement, and it is a requirement of the design rather than a later refinement.
No. The dropzone on this page reads the image in your browser and nothing is uploaded anywhere. When the extraction layer ships, an analysed image will be stored with your analysis and deletable with it.
The first read is free with no account — on a symbol, which is what works today. Beyond that, an account keeps your analyses, journals your decisions and tracks how your own calibration develops over time.