01
Scope and fit
We decide where Vision earns its place in your system, and where a simpler tool wins. No resume-driven architecture.
★ VisionLLM Models8-week engagement
VISION · DOCUMENTS, IMAGES, SCREENSFour jobs that share one word.
Reading a document, generating an image, grounding a click on a screen, and segmenting an object are separate problems with separate leaders. The frontier text models now do most of the reading, specialist parsers still win on dense tables, and the cheapest image model has the widest feature set. We route per job rather than per vendor.
Vision APIOCR engineEval pipelines
Cycle
8 weeks · fixed price
Stack
Vision + document AI
Output
Production code + eval suite
Handoff
Full source ownership
[THE SHORT VERSION]
Half the confusion is a team choosing the wrong family.
Generation models make pictures and cannot read them. Vision-language models read and cannot draw. Document parsers extract structure and cannot reason about it. Teams routinely pick an image model for a comprehension job because both are called vision, then spend a sprint discovering the mistake. We settle which of the four jobs you actually have before anyone opens a pricing page.
When it fits
- Extracting structure from invoices, filings, forms, and contracts
- Reading dense charts and tables nested inside PDFs
- Agents that operate a real browser or desktop interface
- Generating and editing product, marketing, or catalogue imagery
- Segmentation and object boundaries in an image pipeline
When it does not
- Documents with a fixed layout that a deterministic parser already handles
- Anything where a wrong extraction reaches a customer without a review step
[HOW WE BUILD IT]
How we build with Vision.
02
Build on a tested foundation
We integrate Vision against a foundation we trust: typed code, CI, and observability from the first commit. Boring infrastructure, modern surface.
03
Eval before launch
An eval suite proves the build behaves before it reaches a user. We measure, then ship.
04
Handoff with ownership
Your team gets the code, the tests, and a runbook. No lock-in to us or to a vendor framework.
[WHAT YOU GET]
What the engagement leaves behind.
Senior
Engineers who have shipped this before
100%
Source ownership at handoff
Eval-first
Tested before it ships
0
Framework lock-in
[MODELS + VENDORS]
Four jobs, four different leaders.
Document understanding, image generation, screen grounding, and segmentation are separate problems that share a word. No vendor leads all four, and the frontier text models now do a lot of the reading, so the routing decision matters more here than the model does. We integrate them directly behind one abstraction and pick per job.
FlagshipOpenAI
Apr 2026GPT-Image-2
gpt-image-2
Image generation and editing
Image out
$30 / 1M
Image in
$8 / 1M
- Token-priced rather than per-image, so cost tracks resolution and detail
- Generation, edits, and inpainting, with a batch path at half rates
- Image input cached at $2 per million, a quarter of the uncached rate
Read the technical brief
FlagshipGoogle
2026Nano Banana Pro
gemini-3-pro-image
Image generation and editing, top tier
1K / 2K image
$0.134
4K image
$0.24
- Per-image pricing you can quote to a client before you build
- Up to six object reference images, semantic-mask editing, 1K to 4K output
- Legible stylised text, which is the failure mode that used to rule generation out
Read the technical brief
CurrentGoogle
2026Nano Banana 2
gemini-3.1-flash-image
Image generation, high volume
1K image
$0.067
4K image
$0.151
- Half the price of Pro at 1K, with a wider reference-image surface
- Up to 10 object, 4 character, and 3 style references, plus ten aspect ratios
- 512px output as well, which Pro does not offer
In the routing table · brief not yet published
PreviewGoogle
2026Gemini 3.1 Pro
gemini-3.1-pro-preview
Document understanding at volume
Billing
Per token, preview
- Native PDF input, so no image-conversion stage in your pipeline
- Leads OCR and visual question answering on the Nanonets IDP leaderboard
- Handles sparse tables at 94%, the case that breaks most document pipelines
In the routing table · brief not yet published
FlagshipOpenAI
3 Sep 2026GPT-6 Astra
gpt-6-astra
Screen grounding and computer use
Input
$10 / 1M
Output
$50 / 1M
- 92.7% on ScreenSpot-Pro against 76.9% for GPT-5.6 Sol
- The model to reach for when an agent has to click a real interface
- Image in, text out. It reads screens, it does not draw them
In the routing table · brief not yet published
FlagshipAnthropic
1 Sep 2026Claude Fable 5.1
claude-fable-5-1
Dense documents and charts
Input
$10 / 1M
Output
$50 / 1M
- Built for dense charts, filings, and tables nested inside PDFs
- Crop-and-zoom tools on charts, which is how a human reads them too
- 1M context, so a long filing and its exhibits fit in one pass
In the routing table · brief not yet published
CurrentOpen weights
2026PaddleOCR VL 1.5
PaddleOCR-VL-1.5
Document parsing specialist
Billing
Self-hosted
- Tops OmniDocBench at 94.37 overall, ahead of every general model on that test
- Self-hostable, which settles the question when documents cannot leave your network
- A parser, not a reasoner. Pair it with an LLM for anything beyond structure
In the routing table · brief not yet published
CurrentOpen weights
2026Qwen3-VL
Qwen3-VL
General vision language model
Billing
Self-hosted
- Competitive with proprietary frontier models across multimodal benchmarks
- Covers grounding, video understanding, OCR, and document comprehension
- The open-weight default when you need vision inside your own boundary
In the routing table · brief not yet published
CurrentMeta
2026SAM 3
sam-3
Segmentation
Billing
Self-hosted
- Segment anything: masks and object boundaries rather than descriptions
- The right tool when you need pixels back, not prose
- Pairs with a VLM, which names what SAM has isolated
In the routing table · brief not yet published
[METHODOLOGY · K-FRAMEWORK]Integrated through the
Integrated through the
K-Framework.
Every model we integrate runs through the same operating system. Three pillars, sixteen layers, one Compound Growth Loop. The methodology that keeps AI work from rotting after the first ship.
Read the K-Framework01
Foundations
Direct API integration with the model. No LangChain, no orchestration vendor, no agent framework built on quicksand. Typed contracts, the same way we wire up Postgres.
02
Amplification
An eval suite built from your real tasks gates every prompt and model change. Quality is measured before it ships, not vibed in a demo.
03
Judgment
Governance, audit, and oversight wired in from day one. Who called what, with which prompt version, at what cost. Your auditors get answers, not screenshots.
[OBSERVABILITY]
Observability your team can read.
A model in production without observability is roulette. We instrument every integration so engineering and finance can see the same numbers, and so a regression at 3am surfaces before a customer opens a ticket.
Instrumented
Cost per call
Tokens in, tokens out, dollars spent. Sliced by feature, tenant, and route. Budgets enforced where it matters.
Instrumented
Latency p50 / p95 / p99
Real distributions, not averages. We know which routes are slow, and why.
Instrumented
Eval pass rates
The same eval suite that gates a release runs continuously in production. A regression on real traffic surfaces fast.
Instrumented
Prompt + completion logs
PII scrubbed at the proxy, shipped to your SIEM. Retention controls match your compliance window.
Dashboards your team owns, not ours. At handoff you get the queries, the alerts, and the runbook. We are not in the path to read your metrics.
[RELATED]
Worth a look next.
APPLIED K-FRAMEWORKBring the problem.
Bring the problem.
We’ll bring the build.
Senior engineers, eval suite at handoff, full source ownership. Sprint, program, or ongoing. We shape the engagement to the work.