SolaryxDev: Intelligent systems, engineered to run your business.
An engineering firm for autonomous agents, private models and the systems beneath them, working with companies worldwide.
ScrollEngineering firm · Worldwide
Agents you can audit, on infrastructure you control.
SolaryxDev takes on the hard end of applied intelligence. We build agents that carry multi-step work through live systems, and the models behind them can run privately, on hardware you choose. Both are kept correct after launch by evaluation sets, reconciled data, cost budgets per run, drift alarms and upgrades that ship only once they pass.
We run what we build, and answer for it afterwards.
- Fixed price against a signed scope
- 12-month warranty on systems we host
- Operated to the GDPR, hosted in Germany by default
- Source code yours on full payment
- Your data trains no model others can use
- No third-party trackers on this site
Capabilities
Four disciplines, engineered as one system.
Fifteen capabilities, from the first agent to the platform it runs on.
Autonomous systems5 services
- Autonomous agentsMulti-step work carried across your systems through narrow tools, behind a gate no prompt can pass.
- Knowledge & retrieval (RAG)Retrieval over your own documents, with every answer cited to its file and page.
- Document intelligenceSupplier paperwork read into structured records, with the straight-through rate measured on your files.
- Computer visionCameras, scans and forms turned into records, every read carrying a confidence score.
- Voice & conversation systemsCalls and chats filed against the right customer, with the promises in them turned into tasks.
Models & assurance3 services
- Private & sovereign modelsOpen-weight models on hardware you control, or hosted for you in Germany and operated to the GDPR.
- Evaluation & assuranceEvaluation sets, refusal cases and a release gate that every model change must pass.
- Model operations (LLMOps)Routing, cost budgets per run, drift alarms and tested upgrades for models in production.
Data & platforms4 services
- Data foundationsRecords cleaned, migrated and reconciled before any model or system is allowed to rely on them.
- Integration & orchestrationThe systems you already run, connected so one figure travels through every one of them.
- Custom platformsThe system nobody sells off the shelf, engineered to be handed over with its source.
- Software modernizationAgeing or half-built software put under test, then refactored, finished or replaced in slices.
Strategy & operations3 services
- Strategy & advisoryWhere a model will pay for itself in your operation, and where it will not.
- Team enablementHands-on sessions inside your own systems, including when to leave the model out.
- Managed hosting & supportHosting, monitoring and restore-tested backups, run by the team that built the system.
By industry
One engineering standard, across ten industries.
Search by industry, or by the problem in front of you.
- Clinics & healthcareServicesBilling, reminders and front-desk calls, all reading from the diary you already keep.What we buildBilling that matches the visitReminders and recallsFront-desk calls, summarised
- Events & venuesServicesTickets that cannot oversell, and a settlement where sold, scanned and paid agree on the night.What we buildSettlement and reportingPromoter and venue splitsPostponements and refunds
- Professional servicesServicesProposal, e-signature, milestone billing and collections generated from one record, with every change of scope signed before it is billed.What we buildProposals from a recordMilestone, retainer and time billingChange orders
- Real estate & property managementServicesLeases signed online, rent and cheque dates chased on a rule, and maintenance requests tracked until their cost lands on the unit.What we buildLeases signed in the browserMaintenance tracked to costCalls summarised by unit
- Accounting & audit firmsServicesThe documents clients send you read into fields for your staff to review, and an archive of working papers that answers with the file and page cited.What we buildCapture into fieldsValidation before postingAn answerable archive
- Trading & distributionIndustry & supply chainInvoices, delivery notes and customs papers matched to the order, with freight and duty carried into the cost of every item.What we buildSupplier document captureThree-way matchingLanded cost
- Logistics & freightIndustry & supply chainBills of lading, packing lists and carrier invoices read into one record, with every disagreement queued for a person.What we buildShipment document captureCross-document checksProof of delivery on a phone
- Construction & contractingIndustry & supply chainEvery variation priced before it is billed, and every retention tracked to its release, on one record per contract.What we buildSubcontractor papers read and chasedSigned, numbered variationsClaims from measured work
- Manufacturing & industrialIndustry & supply chainBatch-traced stock and work orders on one record, with cameras counting and checking the line as it runs.What we buildCountingPresence and labellingPacking and visible defects
- Retail & e-commerceIndustry & supply chainOrder, stock and returns questions settled on WhatsApp and web chat, with payments and stock reconciled behind every answer.What we buildPayments against ordersStock across channelsReturns and refunds
Not listed? Most trades run on the same record. Tell us about yours.
Intelligence
The frontier, measured. Seven charts, each read at its source.
What the systems can now do, what that costs, where the money and the use are going, and what peer-reviewed studies find when people work with them. Every figure comes from its primary source, and each chart links to it.
Fig. 1Capability
4.2monthsfor the length of task a frontier model can finish on its own to double, on METR's fit to frontier releases since 2023.
The work an agent can finish alone keeps getting longer.
Length of task, in the time it takes a skilled person, that a model completes half the time, by release date. Frontier models in blue. Above 16 hours METR calls its measurements unreliable and marks no model as frontier, so the point there is open and grey.
Source: METR, Time horizons of frontier models, Time Horizon 1.1, updated 8 May 2026. The trend line is fitted to the frontier models since 2023, as METR fits it.
Data table
| Model | Released | Task length, 50% success | Interval, low | Interval, high | Frontier model |
|---|---|---|---|---|---|
| GPT-2 | Feb 2019 | 3 s | 1 s | 9 s | Yes |
| GPT-3 | May 2020 | 9 s | 6 s | 13 s | Yes |
| GPT-3.5 | Mar 2022 | 36 s | 15 s | 1.1 min | Yes |
| GPT-4 | Mar 2023 | 4.0 min | 1.9 min | 8.0 min | Yes |
| GPT-4 Nov '23 | Nov 2023 | 4.0 min | 1.9 min | 8.4 min | Yes |
| Claude 3 Opus | Mar 2024 | 4.0 min | 1.7 min | 8.8 min | No |
| GPT-4 Turbo | Apr 2024 | 3.7 min | 2.0 min | 6.7 min | No |
| GPT-4o | May 2024 | 7.0 min | 4.0 min | 13 min | Yes |
| Claude 3.5 Sonnet (Old) | Jun 2024 | 11 min | 5.5 min | 22 min | Yes |
| o1-preview | Sep 2024 | 20 min | 12 min | 33 min | Yes |
| Claude 3.5 Sonnet (New) | Oct 2024 | 21 min | 10 min | 41 min | Yes |
| o1 | Dec 2024 | 39 min | 21 min | 1.1 h | Yes |
| Claude 3.7 Sonnet | Feb 2025 | 1.0 h | 33 min | 1.7 h | Yes |
| o3 | Apr 2025 | 2.0 h | 1.2 h | 3.2 h | Yes |
| Claude Opus 4 | May 2025 | 1.7 h | 1.0 h | 2.7 h | No |
| Claude Opus 4.1 | Aug 2025 | 1.7 h | 59 min | 2.7 h | No |
| GPT-5 | Aug 2025 | 3.4 h | 1.9 h | 6.8 h | Yes |
| Gemini 3 Pro | Nov 2025 | 3.7 h | 2.3 h | 6.3 h | Yes |
| GPT-5.1-Codex-Max | Nov 2025 | 3.7 h | 2.2 h | 6.6 h | No |
| Claude Opus 4.5 | Nov 2025 | 4.9 h | 2.7 h | 10.4 h | Yes |
| GPT-5.2 (high) | Dec 2025 | 5.9 h | 3.3 h | 13.6 h | Yes |
| Claude Opus 4.6 | Feb 2026 | 12.0 h | 5.3 h | 60.6 h | Yes |
| GPT-5.3-Codex (high) | Feb 2026 | 5.8 h | 3.2 h | 13.6 h | No |
| Gemini 3.1 Pro | Feb 2026 | 6.4 h | 3.9 h | 11.6 h | No |
| GPT-5.4 (xhigh) | Mar 2026 | 5.7 h | 3.1 h | 12.8 h | No |
| Claude Mythos Preview (early) | Apr 2026 | likely at least 16 h | 8.5 h | 55.1 h | Not marked, above 16 h |
Fig. 2Agents
90%of OSWorld-Verified computer tasks completed by the best agent on the leaderboard's All tab, July 2026. People completed 72% of the original tasks.
On everyday computer work, the best agents now pass the human baseline.
Best score to date on OSWorld: real tasks in browsers, office software and the desktop. Self-reported results, and OSWorld-Verified, the revised task set run by the OSWorld team, as each system's mean over its runs. Best-of-N entries are left out.
Source: OSWorld, XLANG Lab, leaderboard and results files, the OSWorld-Verified leaderboard's All tab (its default tab, Foundation E2E GUI, is led by 85.96%). The human baseline, 72.36%, was measured on the original task set.
Data table
| System | Date | Self-reported | OSWorld-Verified, mean of runs | Runs |
|---|---|---|---|---|
| GPT-4, accessibility tree only | Mar 2024 | 12.24% | ||
| Agent S with GPT-4o | Oct 2024 | 20.58% | ||
| learn-by-interact with Claude 3.5 Sonnet, accessibility tree only | Oct 2024 | 22.5% | ||
| OSCAR with GPT-4o, with gold reward | Oct 2024 | 24.5% | ||
| UI-TARS-72B-DPO | Jan 2025 | 24.6% | ||
| OpenAI CUA (GPT-4o) | Jan 2025 | 38.1% | ||
| UI-TARS-1.5 | Apr 2025 | 42.5% | ||
| OpenAI CUA (o3) | May 2025 | 42.9% | ||
| GTA1 with o3 | Jul 2025 | 45.2% | ||
| GTA1 with o3 | Jul 2025 | 53.1% | 1 run | |
| Agent S2.5 with o3 | Jul 2025 | 56.0% | 1 run | |
| CoACT-1 | Aug 2025 | 60.76% | 1 run | |
| GTA1 with GPT-5 | Oct 2025 | 63.41% | 1 run | |
| AskUI VisionAgent | Nov 2025 | 66.2% | ||
| GBOX Agent | Nov 2025 | 64.22% | 1 run | |
| UiPath Screen Agent with Claude Opus 4.5 | Dec 2025 | 67.14% | 1 run | |
| HIPPO Agent with Claude Opus 4.5 | Feb 2026 | 74.48% | 1 run | |
| OpenAPA with Gemini 3.1 Pro | Apr 2026 | 78.34% | 1 run | |
| Holo3-35B-A3B | Apr 2026 | 80.36% | mean of 2 runs, ±2.20 | |
| Pointer Agent with Claude Opus 4.7 | May 2026 | 83.64% | 1 run | |
| Intelligence-Indeed Agent | Jul 2026 | 90.19% | 1 run |
Fig. 3Cost
13×cheaper each year for the same result, about 47% a quarter since 2023.
The same result gets cheaper faster than any major technology before it.
Lowest cost per question, in US dollars, to reach guessing-adjusted scores of 25% and 75% on GPQA Diamond, a graduate-level science test. Guessing-adjusted means 0% is what guessing earns on its four-way multiple choice, so 75% is 81.25% of questions right. Each step down is a cheaper model. Log scale.
Source: Epoch AI, The plunging price of thought, 22 September 2026, figure 2.
Data table
| Model | Released | Cost per question, 25% guessing-adjusted score | Cost per question, 75% guessing-adjusted score |
|---|---|---|---|
| GPT-4 Turbo (Nov 2023) | Nov 2023 | $0.0197 | |
| Gemini 1.5 Pro | Feb 2024 | $0.00194 | |
| WizardLM-2 8x22B | Apr 2024 | $0.000623 | |
| Llama 3.1-70B | Jul 2024 | $0.000209 | |
| Gemini 1.5 Flash | Sep 2024 | $0.000145 | |
| Phi-4 | Dec 2024 | $0.000108 | |
| Mistral Small 3 | Jan 2025 | $0.000089 | |
| o3 (high) | Jan 2025 | $0.299 | |
| Gemini 2.5 Pro (Jun 2025) | Mar 2025 | $0.0891 | |
| Gemini 2.5 Pro (Jun 2025) | Jun 2025 | $0.0129 | |
| gpt-oss-20b (low) | Aug 2025 | $0.000085 | |
| DeepSeek-V3.2 | Dec 2025 | $0.00882 | |
| Qwen3.5 397B-A17B | Feb 2026 | $0.0071 | |
| Qwen 3.5 Flash (hosted 35B-A3B) | Feb 2026 | $0.00218 | |
| MiniMax-M3 | Jun 2026 | $0.00121 | |
| GPT-5.6 Luna (low) | Jul 2026 | $0.000412 |
Fig. 4Capital
USD 582bnof corporate investment worldwide in 2025, up 130% on 2024.
Corporate investment more than doubled in a year.
Global corporate investment in the field, by kind, billions of US dollars, 2013 to 2025.
Source: Stanford HAI, AI Index Report 2026, figure 4.2.1, from Quid data.
Data table
| Year | Private investment | Mergers and acquisitions | Minority stakes | Public offerings | Total |
|---|---|---|---|---|---|
| 2013 | 5.17 | 5.92 | 1.93 | 1.55 | 14.57 |
| 2014 | 9.56 | 6.69 | 0.34 | 2.45 | 19.04 |
| 2015 | 13.35 | 8.85 | 1.18 | 2.05 | 25.43 |
| 2016 | 17.13 | 13.05 | 1.61 | 2.03 | 33.82 |
| 2017 | 25.72 | 24.68 | 1.29 | 2.03 | 53.72 |
| 2018 | 43.10 | 21.89 | 2.01 | 12.62 | 79.62 |
| 2019 | 58.18 | 36.43 | 2.15 | 6.51 | 103.27 |
| 2020 | 73.79 | 39.83 | 88.19 | 20.06 | 221.87 |
| 2021 | 145.40 | 175.36 | 2.65 | 37.32 | 360.73 |
| 2022 | 113.01 | 121.39 | 8.88 | 9.97 | 253.25 |
| 2023 | 104.34 | 82.26 | 8.98 | 5.42 | 201.00 |
| 2024 | 151.48 | 92.19 | 4.49 | 4.86 | 253.02 |
| 2025 | 344.66 | 214.44 | 5.93 | 16.66 | 581.69 |
Fig. 5Adoption
88%of respondents' organisations use AI in at least one business function, 2025. In 2017 it was 20%.
Nearly nine in ten respondents say their organisation now puts models to work.
Share of survey respondents whose organisation uses AI in at least one business function, and generative models in particular. Self-reported, so directional.
Source: Stanford HAI, AI Index Report 2026, figure 4.3.1, from McKinsey’s global surveys.
Data table
| Year | Any use | Generative models |
|---|---|---|
| 2017 | 20% | not reported |
| 2018 | 47% | not reported |
| 2019 | 58% | not reported |
| 2020 | 50% | not reported |
| 2021 | 56% | not reported |
| 2022 | 50% | not reported |
| 2023 | 55% | 33% |
| 2024 | 78% | 71% |
| 2025 | 88% | 79% |
Fig. 6Evidence
26%more tasks completed by 4,867 developers given a coding assistant, across three randomised trials.
Across four studies, the gains are real, and uneven.
Change for people working with a model against a comparison group without one, in peer-reviewed studies. Faster is shown as a gain. One task was chosen to sit beyond what the model does well. Where a study finds no significant change, no bar is drawn.
Sources: Cui et al., Management Science, 2026; Dell'Acqua et al., Organization Science, 2026; Noy and Zhang, Science, 2023; Brynjolfsson, Li and Raymond, QJE, 2025.
Data table
| Measure | Change | Study |
|---|---|---|
| Tasks completed | +26.1% | Cui, Demirer, Jaffe, Musolff, Peng and Salz, Management Science, 2026; 4,867 developers in three randomised trials |
| Quality of work | +32% | Dell'Acqua et al., Organization Science, 2026; 758 consultants in a randomised trial |
| Speed | 25.1% faster | Dell'Acqua et al., Organization Science, 2026; 758 consultants in a randomised trial |
| Tasks completed | +12.2% | Dell'Acqua et al., Organization Science, 2026; 758 consultants in a randomised trial |
| Right answers, beyond the frontier | −19 points | Dell'Acqua et al., Organization Science, 2026; 758 consultants in a randomised trial |
| Time taken | 40% less time | Noy and Zhang, Science, 2023; 453 professionals in a randomised trial |
| Quality | +18% | Noy and Zhang, Science, 2023; 453 professionals in a randomised trial |
| Least-skilled fifth of agents | +36% | Brynjolfsson, Li and Raymond, The Quarterly Journal of Economics, 2025; 5,172 agents, access rolled out in stages |
| All agents | +15% | Brynjolfsson, Li and Raymond, The Quarterly Journal of Economics, 2025; 5,172 agents, access rolled out in stages |
| Most-skilled fifth of agents | no significant change | Brynjolfsson, Li and Raymond, The Quarterly Journal of Economics, 2025; 5,172 agents, access rolled out in stages |
Fig. 7Returns
67%of respondents report a revenue increase from using analytical models in marketing and sales, 2025.
Where respondents report the returns.
Respondents reporting a revenue increase or a cost decrease from analytical models, by business function, 2025. Self-reported.
Source: Stanford HAI, AI Index Report 2026, figure 4.3.4, from McKinsey’s 2025 survey.
Data table
| Function | Revenue increase | Cost decrease |
|---|---|---|
| Marketing and sales | 67% | 49% |
| Strategy and corporate finance | 65% | 53% |
| Product or service development | 62% | 47% |
| Supply chain management | 59% | 49% |
| Service operations | 57% | 51% |
| Software engineering | 57% | 56% |
| IT | 52% | 54% |
| Manufacturing | not reported | 56% |
| Human resources | not reported | 51% |
| Risk | not reported | 45% |
| Knowledge management | not reported | 44% |
In practice
Autonomy, with a person at every irreversible step.
Spending money, deleting data and first contact with a customer pass through a gate that sits outside the model. No prompt can argue its way past it.
-
Agents that act inside your systems
They read the record, work in your ERP, calendar and inbox, and log every step. Irreversible actions wait for a person.
Return matched to order and payment
Refund drafted
Refund: held for a person
-
Documents and cameras, read into records
Scans, forms and camera frames become records, each read with a confidence. Anything under the threshold queues for review and posts nothing.
Carton count0.97posted
Label0.64queued for review
-
Calls, filed and acted on
Transcribed in English and Arabic and summarised against the right customer. The promises made on the call become tasks.
Call summarised on the customer's record
Task: send the revised quote by Thursday
Task: book the site survey
-
Answers you can check
A private assistant that answers from your own documents and cites the file and page. Under its confidence threshold it says it does not know, and it can run on hardware you control.
QWhat did we agree on the SLA?
AFour hours, business days, capped.
MSA-2026-04.pdf · p. 3
Worked examples. Not client data.
Each system ships with the test set that proves it, and no model upgrade goes live until it passes. Models are chosen per job and named in your proposal, open-weight models among them.
Models and technologies we build with. No partner, reseller or referral tie to any of them.
- Anthropic Claude
- OpenAI GPT
- Google Gemini
- Meta Llama
- Mistral
- Qwen
- llama.cpp
- PostgreSQL
Data & platforms
Connected to what you run, reconciled before anything acts.
Data work comes first here, because an agent acting on records that disagree will act on the wrong one. Systems that work stay where they are, and one is replaced only when replacing it pays for itself.
Connect
Your systems reached through their API, a database view or a scheduled export.
Map
Every field traced from its old source to its new home, and written down.
Reconcile
Totals agreed record by record, with an exception list for anything that does not.
Act
Agents act only through narrow, typed tools, each scoped to one job.
Book
One figure from what is shown to what is charged. Money checks fail closed.
Integration & orchestration
Enquiry to quote, contract to invoice and payment matching, carried across the systems you run.Custom platforms
Internal operations tools, approval flows and middleware, built for the process that sets you apart.Software modernization
Today's behaviour pinned down by tests before a line changes, then rebuilt one slice at a time.
Security & data protection
Your data stays where the contract puts it.
We run every system to the GDPR, and none of the controls below is an add-on: each one is there from the first release.
Hosted inside the EU
Germany by default, or wherever your contract says, your own servers included. Personal data from the EU crosses a border only with a transfer mechanism the GDPR recognises.
Processed on your instructions
You stay the controller. We process only on your documented instructions, report any breach without undue delay, and delete or return the data when the agreement ends.
One client, one boundary
Each client's system is walled off from every other, and a record is served only to a person allowed to see it.
Nothing learned from your data
Your documents inform an answer and teach the model nothing. No model that anyone else can use is trained on them.
Proof
Evidence you can inspect.
Client names stay private until each client agrees in writing to be named, so there are no logos here yet. What we can show is how the work is done, with figures you can hold us to.
The release gate
Our own platform's last full release, on 10 September 2026, cleared every one of its 2,584 automated checks, and each of the 43 client documents it renders came out complete. Nothing was deleted and nothing had to be reverted.
Releases that can go back
Every release runs next to the one before it and goes live only after it answers correctly, and the earlier one is kept running.
Our own company runs on it
We keep our own accounts, contracts and client email on the same stack. Our own agreements are signed on our own platform, each copy kept with its signer record.
How we engage
One fixed price from first call to production.
You approve one number against a written scope before anything starts. A change of scope arrives as a written quote with its fee and dates, and waits for your approval in writing or in your portal.
Call
Thirty minutes on what the business does and what done would look like. No deck.
Proposal
Scope, milestones and one total, read and signed in your browser.
Build
Milestone by milestone, visible in your portal as it happens.
Launch
Beside the old version, which stays in place until the new one proves itself.
Warranty
12 months while we host it, then the Premium Subscription or a clean handover.
Four questions buyers ask first.
- What does it cost, all of it?
- One fixed total for the build, set out in the proposal with its payment schedule: a deposit on signature, then milestones. Hosting and the optional Premium Subscription after launch are priced in the same proposal. Nothing is invoiced that is not on the signed schedule or a change you approved in writing.
- How long does it take?
- The dates are in the signed proposal, one for each milestone. If we are waiting on your access, content or a decision, the dates move by the length of the wait, as the contract sets out.
- What happens when something fails?
- A defect in what we built is fixed at no cost during the warranty. A release that misbehaves is rolled back to the version before it.
- Who answers when it breaks?
- A person. You raise it in your client portal, where your requests sit beside your documents and invoices, and during the warranty a defect in what we built is corrected at no cost. The Premium Subscription adds target response times by severity, set out in its own agreement.
Engagement models
Anywhere you operate, on one set of terms.
We work with companies worldwide, remotely. The person who scopes your system also writes it, so no handover sits between the proposal and the code.
Project
A scoped build with milestones and one total, signed online and warranted after launch.
Base Platform Care
The hosting itself, from the system's second year, on a yearly fee set in your proposal.
Premium Subscription
Ongoing changes and priority on anything that stops work. While it runs, the warranty keeps extending, up to 24 months.
Handover
We build it, hand it over, then support it on servers you run.
Commitments
Commitments the contract holds us to.
| Commitment | In one line |
|---|---|
| Price and payment | A fixed total on a written scope, milestone payments, and a written quote before any change. |
| Warranty | 12 months on our hosting, as many as 24 with the Premium Subscription, 90 days on your own servers. |
| Ownership | What you have paid for is yours, source code included, and no clause locks you in. |
| Leaving | Move to your own servers or another provider whenever you choose. We hand over your data for the move. |
| Independence | No partner, reseller or referral tie to any provider we recommend. |
Contact
Tell us the problem.
Answered by the engineer who would build it.
Thanks, your message is with us.
A person reads it and replies.