Explainer video (about 9 min)
Other language versions: Japanese / English / Simplified Chinese / Playlist
0. Summary—what to build, what to adopt, and what people decide
In one sentence, it is "a company that handles on-site electricity and web encryption with the same hands". The fact that the hand that lays the DC wiring for solar power plants and the hand that writes code to seal estimates with a key management service are the same is what determines how this company is built.
What we build and operate ourselves are the points of contact with customers (consultation desk, estimates, email), the defenses of our own site (headers, entry limits, signed links), the delivery mechanisms (translation, video, data publication), calculation tools (30-minute values, reflected light, string design, survey models), and the servers that speak directly with field equipment (OCPP, Modbus). What we take in from manufacturers are product-side technologies such as SolarEdge optimizers and SafeDC, Tesla Wall Connector, Huawei SmartLogger, and OCPP-compatible chargers. What people decide are the amount, whether construction is possible, the full application of the advertising budget, and the final judgment on published text.
1. Consultation desk—answers in 14 languages, with a gatekeeper stopping the numbers in-house developed
The site’s consultation desk handles inquiries about EV charger installation, self-consumption solar, O&M, and recruitment. It identifies the language entered and responds in 14 languages—Japanese, English, Chinese (Simplified and Traditional), Spanish, French, German, Swedish, Vietnamese, Korean, Thai, Hindi, Arabic, and Hebrew—based on the same knowledge and the same rules.
A single reply is not written all at once. First, we plan what to answer and what to ask as internal structured data; next, we write a short body; third, we look for problems through deterministic checks and review by another AI; and if anything is found, we revise it. With this plan, body, review, and revision process, we call AI up to four times per reply. The first reply is kept short, and in principle we ask only one question.
Gatekeeper for numbers and model numbers. It extracts amounts, percentages, and model numbers in replies (such as Tesla, OCPP, SolarEdge, and JC-STAR), and detects, without using AI, assertions not found in the internally verified knowledge or agreement with a customer’s incorrect premise by means of regular expressions. If anything is found, the reply is sent back. By design, we do not allow AI to make up “plausible numbers” and send them.
We also accept photos. If you send photos of the distribution board, parking location, and exterior wall, AI will read the images and narrow down the points to confirm during the site visit. What we do here is only organizing the estimate conditions and narrowing down the issues, and the amount is not finalized. The amount and whether installation is possible are determined by the person in charge after on-site confirmation (for details, see "How far can same-day EV charger installation and AI preliminary estimates go?").
Consultations with contact details are notified internally, and those notifications include a summary of the request, an analysis of the situation, and a draft reply. By the time the staff member opens the email, they already have a reply ready. If generating the draft takes time, we switch to a standard template, and the notification is not delayed.
2. AI that checks itself—consultation every morning at 8:20 and reflection once a week in-house developed
The quality of the inquiry desk cannot be maintained by people looking in only from time to time. At I-S3, we place three layers of monitoring.
- Every morning at 8:20, it starts an actual consultation. It begins consultations with the live desk in four scenarios and checks for empty responses, server errors, exposure of internal terms, or topic mix-ups. The same check also runs immediately after a new version is deployed, and notifications are sent only when it fails.
- Every Friday, it reflects on 12 questions. It applies 12 questions to the most recent conversation records, such as whether customer anxiety was reduced, whether there was really only one question that needed to be asked, and whether internal matters leaked, and reports them divided into items that need confirmation and items that can be safely reflected. When the rate of obtaining contact information or the rate of ending with the next action drops, those are also treated as “action required.” These results are fed back into the response rules document and take effect from the next reply.
- Outage monitoring every 30 minutes, and attack analysis every Monday. It rechecks AI service outages and authentication errors every 30 minutes, and classifies and reports attack strings sent to the consultation desk—prompt injection such as “ignore previous instructions,” script injection, and SQL injection—every week.
Reading logs later (weekly), checking incident records (every 30 minutes), and actually starting consultations (every morning). With three perspectives that differ in time axis and method, we capture forms of breakdown that cannot be seen with just one.
3. Index of field knowledge—RAG without a vector database in-house developed
The knowledge used by the consultation desk to answer is in the company’s Google Drive documents and spreadsheets. These are ingested every hour and passed to prompts, separated into public chat, solar, recruitment, and internal use only. If you add an internal-only classification at the top of an internal memo, that content will not be mixed into public chat. While using the same knowledge base for both public and internal use, the boundary is managed by a single line on the document side.
We do not use a general vector database for retrieving related knowledge. We split the imported documents, and from each fragment we use AI to extract six types of "entities"—products, equipment, regions, systems, organizations, and procedures—and their aliases to build an index. At answer time, we do not call AI; we retrieve related fragments only by text matching against the question. This avoids increasing latency and cost at answer time, and lets people trace why each fragment was retrieved. If the index becomes outdated, we automatically switch to the conventional method and re-extract only the changed parts of the documents.
4. From quotation to order—sealing with passkey signatures and key management services in-house developed + Google Cloud
The quotation is sent as a PDF, and the customer reads the contents on a dedicated page before confirming the order. There are two ways to confirm: handwritten signature or device identity verification using fingerprint or facial authentication on a smartphone. The latter uses WebAuthn (FIDO2) and again requests a signature from the registered device (such as ECDSA P-256) at the moment the order button is pressed. An order is not completed by entering a name alone. The proof of registration has a short validity period and cannot be reused.
When the signature is established, the hash of the documents at that point, the hash of the signature image, the proof of identity, and the viewing record are combined into one normalized data set, and signed with an asymmetric key in Cloud KMS (Google's key management service). The application itself does not hold a private key and only requests signatures. The signed data and the certificate PDF are stored in a dedicated repository, so later a third party can verify “which document, who, and when consented.” Double clicks and submissions from multiple tabs opened at the same time are allowed through only once by a database transaction.
Links for staff to view the full conversation or the photos sent are issued by signing an unpredictable, time-limited token with HMAC. There is one design commitment: if no signing key is configured, no links are issued at all (fail-closed). The background to this commitment is described in Chapter 12.
5. Email — duplicate delivery paths and do not reply to spoofing in-house development
Estimate copies, receipt replies, and internal notifications are all delivered by email. We do not rely on a single delivery route; if the first route (Amazon SES) cannot send, we switch to the second route (Gmail API), and if both fail, we queue the message and resend every 10 minutes. Delivery health is checked every 30 minutes, and when a route is exhausted we notify the chat tool.
Acknowledgment replies to inquiry emails are created using the same “plan, body, review” flow as the consultation desk, but two checks are applied before sending. One is do not reply to automated emails. Headers carried by out-of-office notices or mailing list messages (such as Auto-Submitted in RFC 3834) are checked to stop replies, and the same header is attached to automated replies sent by us to prevent infinite reply loops in both directions. The other is do not reply to emails whose sender is not authenticated. The authentication results at receipt (SPF, DKIM, DMARC) are read, and if there is no passing result consistent with the sender domain, the auto-reply is stopped and the fact is added to the internal subject line. This is to prevent politely replying to emails that spoof the sender and becoming a relay for spam to third parties.
6. Protect the site — full CSP deployment, zero violations, per-entry limits in-house development
On October 10, 2026, Content-Security-Policy was fully deployed across the site. It is a mechanism that lists the scripts, connection destinations, and frames that the browser may load, and blocks everything else. First, in Report-Only mode for about three hours, we collected violation reports from actual visits and browser crawls of all pages in five languages, corrected the allowed destinations in five categories, and then switched to full deployment. There were zero violation reports after full deployment. The intake for violation reports is maintained in-house, and any parts that may contain personal information are discarded before aggregation.
We keep the setting that allows inline scripts. Many existing pages are written inline, and if we remove them all at once the site would stop. As each page is updated, we move it out, and eventually remove the allowance.
We attach HSTS, nosniff, Referrer-Policy, and Permissions-Policy to all responses, refuse frame embedding for staff links and APIs, and enforce HTTPS on all paths. For the anonymous estimate, visit, diagnosis, and image forms, we set limits by source network unit, and if they are exceeded we return 429 and the number of seconds until retry. Individual IP addresses are not kept in state or logs. Uploaded images are identified by format from the leading bytes rather than the file extension, and on every deployment we verify that static file delivery cannot escape the public folder.
7. Deliver — body text in five languages, video in three languages, reusable data in-house development
The canonical text of the site is Japanese. Each time we deploy, we extract only the text portions of each page and generate translation data in English, Vietnamese, Simplified Chinese, and Traditional Chinese, with a glossary attached (company name, representative’s name, and other specific wording such as repowering). On the browser side, we replace only the text while preserving the page structure. The translation data amounts to about 2,000 files. In translation checks, we automatically discard the intrusion of other writing systems such as Hangul and Cyrillic, as well as incorrect transliterations of personal names. Parts that must not be translated, such as addresses, submission text, and button labels, are marked as excluded from translation and displayed in Japanese as-is.
We apply the same idea to videos, making three versions per article as standard: Japanese, English, and Simplified Chinese. Script generation, narration cleanup, speech synthesis, subtitles, and pre-publication fact checking—including detecting misleading assertions such as "you can directly charge an EV at night with daytime solar" using regular expressions—are all part of one workflow. Article and social media links are tied to the display language, so if you open the article in English, the English video is embedded.
In research articles (economics of dynamic wireless power transfer while driving, cyber ripple effects of distributed power sources, and others), we distribute not only the conclusions but also the Python calculation model, the CSV assumptions, and the results of random trials under CC BY 4.0, describe them in machine-readable form as schema.org Dataset, and include citation format (BibTeX). We believe that the value of primary information lies in being independently checked and cited by third parties. We also distribute llms.txt for AI crawlers, sitemaps, and feeds, and notify search engines of updates with IndexNow on every deployment.
8. Calculate — 30-minute values, reflected light, strings, Monte Carlo in-house development
Non-AI mathematical processing also supports on-site decisions.
- Automatic analysis of 30-minute values. When you paste or upload CSV or Excel demand data downloaded from the utility company, the system tries three reading patterns at the same time—wide format, long format, and transposed—and adopts the most plausible structure, then returns daily statistics and heatmaps. If it cannot be read, it falls back to “a staff member will check it.” The original file is automatically discarded after the retention period expires. This is the first material used to determine the capacity of self-consumption solar PV (how to read 30-minute value data).
- Reflected light simulation. Using the address as coordinates, and from the year, tilt angle, azimuth angle, viewpoint height, and the target's height and distance, it generates annual reflected-light directions and time-stamped diagrams and CSV files. It is a tool for answering questions about reflected light to neighbors before installation.
- I-S3 String Planner. This is a free tool that lets you arrange panels against satellite photos or drawings, decide strings and wiring routes, and complete output in PDF, JSON, and CSV entirely in the browser. There is a real example of a 944-panel layout diagram with 16 in series × 7 to 8 in parallel exported to PDF, and it can be used in 11 languages.
- Model of a research article. For the economics of dynamic wireless power transfer while driving, we ran 100,000 × 2 Monte Carlo simulations; for cyber propagation in distributed energy resources, we ran 20,000 draws of uncertain variables and frequency simulations using the swing equation, and we published Red Team reviews from nine perspectives along with the response table. The dialogue simulator in the article (JavaScript) was verified in seven cases to produce the same results as the research model (Python), and readers can rewrite the assumptions and recalculate.
9. Speaking directly with field devices — OCPP server and Modbus logger in-house development (the specifications are industry standards)
EV chargers are not something you install and then forget. I-S3 wrote its own central system for OCPP (Open Charge Point Protocol) 1.6J / 2.0.1 that stays connected to chargers at all times to handle authorization, transactions, and output allocation, and operates it on Cloud Run (A-Charge). The connection to the charger uses a long-lived WebSocket, and it calculates output allocation across multiple units so as not to exceed the facility's power reception limit. When connecting, it verifies each charger’s token with constant-time comparison, and transaction IDs are generated with random values that do not collide across processes. It was separated from App Engine, which serves the web pages, to avoid losing all charger connections when the web is redeployed.
On the solar PV side, we built Virtual Smart Logger, modeled on the menu structure of Huawei SmartLogger3000. It reads inverter registers via Modbus RTU over RS485, and in environments without hardware it runs with a simulator. Writes such as active power limits and reactive power commands are passed through four controls: an allowlist, value range checks, dry-run by default, and audit logs, enforcing a field procedure of "read and confirm before writing." This is not meant to replace the manufacturer's monitoring equipment, but rather to be a tool for carrying out commissioning and inspections in the field using our own procedures.
10. RightOS — verifying rights without holding personal information in-house development (public product)
RightOS is an API for checking with a QR code whether a person has valid rights. It is intended for use cases such as store queues, facility entry, and organizing pickup areas. The core of the design is not holding personal information: API keys store only hashes, webhook signing keys are derived from a master key and not stored, and inputs that attempt to reach external destinations are blocked. The user-facing screens support 20 languages, the API specification is published in OpenAPI, and an MCP server is also provided for calls from AI agents. Live tests against the production environment are run automatically every day.
11. Distinguishing from manufacturer technology manufacturer technology
Much of the system up to this point becomes useful only when combined with the technologies built into the manufacturers’ products. What does each side handle?
| Manufacturer technology (not developed by I-S3) | the part handled by I-S3 |
|---|---|
| SolarEdge power optimizers, SafeDC (a function that lowers module output to about 1V when stopped), and module-level monitoring | Adoption decisions for buildings where people are active, string design, construction, configuration, measurement and recording of post-shutdown voltage at handover, and separation during O&M (DC safety article) |
| Tesla Wall Connector, and standard AC chargers from various manufacturers (such as Wallbox) | Judgment of distribution boards, main breaker capacity, and wiring routes, same-day installation, on-site estimates, and installation as a Tesla recommended contractor |
| Huawei SmartLogger3000 and inverters, and their Modbus register specifications | On-site commissioning procedures, reading and inspection with Virtual Smart Logger, and manuals for restart and configuration procedures |
| OCPP-compatible chargers (the standard is Open Charge Alliance) | Implementation and operation on the central system (server) side, and output allocation matched to the facility's incoming power limit |
| Demand records in 30-minute intervals provided by the power company | Automatic format detection and parsing, a bridge to calculating self-consumption capacity |
| Google Cloud Cloud KMS, Secret Manager, Firestore, and Cloud Run, and external AI service APIs | A design that does not keep keys ourselves, a design that stops if there is no secret, decisions about which data to send where, and use under contract terms that do not use it for training |
It is true that I-S3 actively adopts new technologies, but adopting them and inventing them are different things. I-S3's technical strength lies in being able to create on its own the series of tasks of selecting, designing, building, configuring, connecting, and operating, as well as the tools used around them.
12. Operating practices — one deployment procedure and four lessons in-house development
Production updates are made with a single script. There are 158 regression tests, and if the security headers defined in the configuration file and the code do not match, the tests fail.
- Verification — matching the deployment project (to prevent recurrence of past incidents)
- Inspection — static route inspection / presence of signing keys / style inspection of published articles
- Generation — translation JSON (4 languages) / FAQ / regional landing pages / case studies / sitemap
- Deployment — gcloud app deploy (fixed project) / deletion of old versions
- Inspection — actually start the consultation desk (4 scenarios) / self-review of version boundaries / delivery health
- Notification — IndexNow / Search Console
This practice was decided by failure.
Lesson 1: If there is no secret, do not run on a weak substitute. Stop. With the key for signed links left unset in production, the code kept running by using a value derived from a publicly available identifier as a substitute key. Links were issued, appeared to work, and no one noticed. After finding it during inspection, we eliminated the substitute, changed the design to not issue links unless there is a key (fail-closed), and intentionally invalidated all links issued in the past. "It works" is not proof that it works safely.
Lesson 2: Count how many things depend on the key of a provider you have stopped using. Even after discontinuing one email delivery provider, two messages to customers (the receipt reply and the estimate copy) remained tied exclusively to that route and became undeliverable. In addition, the "can it be sent" decision for 11 internal notifications depended on whether the discontinued provider's key was present. Before deleting a key, we added a procedure to identify the structure that would stop all at once the moment one key was removed.
Lesson 3: Do not update the same authorization token in two places.The refresh token used for automatic posting to social media was being stored and updated separately by the local environment and the server’s scheduled process. At one point, saving on the local side silently failed because of a name resolution issue, and as a result the server side tried to refresh with an old token, causing the entire authorization to be revoked. We unified the storage location and fixed it so that if saving fails, the reason is always shown. A process that fails silently is the most expensive one.
Lesson 4: A library’s “silent drop” becomes the infrastructure’s “unhealthy.”On the WebSocket server connected to the charger, the library we were using was disconnecting non-WebSocket HTTP requests without any response. The execution platform’s health check judged that as “unhealthy,” which became an incident that rejected all connections. We resolved it by replacing the HTTP parsing part. You need to compare in advance what the platform health check sends and what your server returns.
What the four have in common is that we did not leave an incident as “fixed and done,” but converted it into one of the following: a regression test, a design rule, or a procedure manual. Of the 158 tests, 36 were added in the October 2026 inspection.
13. What we are not doing, and what is still under verification
- AI does not determine the amount.The consultation desk and photo analysis are limited to organizing the conditions and narrowing down the issues. The amount and whether the work can be carried out are decided by a person after an on-site inspection.
- We do not automatically make ad budget changes final.The weekly review proposes improvements, sets an upper limit on the range that can be changed automatically, and requires human confirmation for final approval.
- The content of public chats is not used for training.We use the APIs of external AI services under contractual terms that do not allow the submitted content to be used for model training (Privacy Policy).
- Small language models that run locally are under evaluation.We are testing small models for inquiry classification and translation comparison, but they are not part of the production path. Our design policy is to limit production inference to what runs on the server side.
- The CSP inline allowance remains.We will remove it as the page is updated.
Rather than making what we can do look bigger than it is, we believe it is more useful for both customers and peers to make clear where people are making decisions.
Frequently Asked Questions
Does the AI chat produce the estimate amount?
We do not send it. What the chat does is organize the conditions needed for an estimate and narrow down the points to confirm during the visit. The amount and whether construction is possible are decided by the person in charge after an on-site check. Any numbers or model numbers included in the reply are sent back before transmission if they are not in the internally verified knowledge.
Is what you write in chat used for AI training?
It is not used. We use external AI service APIs for response generation, but under contractual terms that prevent each provider from using the submitted content for model training. Details are described in the Privacy Policy.
What happens on devices where passkeys (fingerprint or face authentication) cannot be used to sign an estimate?
Orders can be confirmed with a handwritten signature. Either a handwritten signature or identity verification on the device is required; entering only a name is not sufficient. The documents and evidence at the time of signing are sealed with a key from the key management service and stored in a form that can be verified later.
Can the mechanisms introduced here be used by other companies as well?
RightOS (rights verification API) and String Planner (solar string design tool) are public and available to anyone. The calculation model and data in the research article are published under CC BY 4.0. Other internal systems are explained in the article in terms of the design approach and the standards adopted. We also accept consultations from those who want to build the same setup in-house.
Are SolarEdge and Tesla technologies developed by I-S3?
That is not correct. SolarEdge SafeDC and optimizers, Tesla Wall Connector, Huawei SmartLogger, and OCPP-compatible chargers are technologies from each manufacturer. What I-S3 is responsible for is their selection, design, construction, configuration, connection, operation, and the in-house tools used around them. They are organized in a table in Chapter 11.
Try the mechanisms introduced in this article for real
The contact desk operates in 14 languages. It accepts not only consultations about EV chargers and solar power, but also requests such as "we want to build the same setup in-house" or "please explain the mechanism of electronic signatures for estimates in detail." Numbers and model numbers are checked by the gatekeeper before being returned, so anything that cannot be answered is returned as "cannot answer."
Open the consultation desk Contact the person in charge directly
Related pages
- About communication and data security (TLS, encryption at rest, signed URLs, handoff to the person in charge)
- Privacy Policy (information obtained in AI chat, terms of use for external AI services)
- How far can same-day installation of EV chargers and AI preliminary estimates go (scope of issue sorting from photos)
- The first asset companies lose in the age of AI is “listening” (an essay on AI adoption and an organization’s ability to listen)
- I-S3 String Planner (solar string design tool)
- Even if the PCS is stopped, solar power generation does not stop (the positioning of SolarEdge SafeDC and DC safety in Japan)
- Cyber Blast Radius——How to connect tens of millions of DERs (an example of model and data disclosure)
- Tateyama Expressway economics of wireless power transfer while driving (an example of Monte Carlo and Red Team review)
- RightOS/A-Charge
- Technical library
This is based on production operations as of October 10, 2026. Standards and product names belong to their respective rights holders. Thresholds, internal configuration values, and the names of keys and identifiers are not listed for safety reasons.