← Back to Home

Frequently Asked Questions

Common questions about Intentioned.tech and social skills training

Moderation & Safety

Why is the default moderation prompt not strict at all?

Modern large language models (LLMs) like Qwen, Gemma, and others already have extensive built-in safety guardrails and content filtering. These models are trained to refuse harmful requests, avoid generating dangerous content, and maintain ethical boundaries by default.

Adding an overly strict moderation layer on top of these already-cautious models creates a problematic "double filtering" effect that severely impedes innocent conversations. Users practicing difficult scenarios—like handling an angry customer, negotiating assertively, or discussing sensitive topics in therapy contexts—would constantly trigger false positives.

Example: A user practicing "How do I tell my boss I'm overwhelmed?" might use words like "stressed," "can't cope," or "breaking down." Overly strict moderation would flag this as a mental health crisis, interrupting legitimate professional communication practice.

Our moderation is designed to catch only the most extreme cases (direct, actionable threats to real individuals) while allowing the natural flow of difficult conversation practice. The underlying LLM handles the rest.

Can I customize the moderation prompt if it's too strict or too lenient?

Yes, absolutely! If you find the default moderation prompt doesn't fit your needs, you can fully customize it in the configuration tool.

How to Customize:
  • Open the Intentioned.tech Configuration Tool
  • Navigate to the 🛡️ Moderation tab
  • Edit the moderation system prompt to match your requirements
  • You can also add/remove keywords from the serious violations list
  • Save your configuration

Some use cases where you might want to adjust moderation:

  • Too strict: You're running a therapy/counseling practice where discussions of mental health topics are expected and legitimate
  • More strict: You're deploying in a K-12 educational setting and want additional safeguards beyond the defaults
  • Domain-specific: You're training customer service agents and need to allow discussion of complaints and frustrations
⚠️ Important: If you make the moderation less strict, ensure you have other safeguards in place (supervision, logging, user agreements) to prevent misuse. You are responsible for the configuration of your deployment.

About the Project

Why does this project appear to be "vibe coded"?

Because it is—and there's a good reason for that.

Intentioned.tech is developed by a single person and an assistant (Ari Kotler and Gianni Latella) as an Honors Thesis project at Pace University. There is no team of developers, no corporate backing, and no venture capital funding. Every line of code, every design decision, and every feature is the work of two individuals balancing this with academic coursework and life responsibilities.

🤖 AI-Assisted Development: Yes, AI tools (including GitHub Copilot, Claude, and other LLMs) are extensively used in the development of Intentioned.tech. However, every AI-generated output is:
  • Carefully reviewed before integration
  • Tested for correctness and security
  • Modified and refined to fit the project's architecture
  • Generated using premium, expensive models (not cheap shortcuts)

The "vibe coded" approach allows rapid iteration and feature development that would be impossible for a solo developer using traditional methods. The alternative would be a project that takes 5 years instead of months, or never ships at all.

Quality Commitment: Despite the unconventional development approach, security and privacy are taken seriously. All data processing happens locally, no conversations are sent to external servers, and the codebase is open for inspection.

If you find bugs or issues, please report them. Solo development means things slip through the cracks—but it also means fixes can be deployed quickly without bureaucratic approval processes.

Demo & Website

Why is the demo password-protected?

The demo is currently private and password-protected for security and quality control reasons:

  • Resource Protection: The demo server runs on limited hardware. Unrestricted access could lead to abuse, overloading, or denial-of-service situations.
  • Quality Assurance: We're still refining the experience. Controlled access allows us to gather feedback from trusted testers before a public launch.
  • Security: Real-time AI applications require careful monitoring. Private access helps us prevent misuse and ensure responsible deployment.
  • Cost Management: Running AI models is expensive. Password protection prevents bots and scrapers from driving up operational costs.
Want Access? If you're interested in testing the demo, please reach out via our contact page or sign up for the mailing list to be notified when public access becomes available.

Where are the demo page links on the home page?

The demo is accessible via the "Launch Demo" button in the navigation bar at the top of the home page. You can also find it in the hero section of the landing page.

⚠️ Demo Availability: The demo runs on a development server and may occasionally be offline for maintenance, updates, or when I'm working on new features. If the demo is down, please try again later or check back the next day.

The demo provides a full experience of Intentioned.tech's conversation practice capabilities, including:

  • Real-time speech-to-text transcription
  • AI-powered conversation partner
  • Natural text-to-speech responses
  • Multiple scenario categories to choose from
  • Language selection (28 EU/EEA/border languages supported)
Tip: For the best experience, use a modern browser (Chrome, Firefox, Edge) with a good microphone. Make sure to allow microphone access when prompted.

Technical Requirements

What language models can I use?

Intentioned.tech supports a wide range of language models through HuggingFace Transformers:

  • Qwen 2.5 (Recommended): 1B, 3B, 7B variants - excellent multilingual support
  • Google Gemma: 2B, 7B, 9B variants - good for general conversation
  • Llama 3: 8B variant - strong English performance
  • Mistral: 7B variant - fast inference
  • GEITje (Dutch): 7B - specialized for Dutch language
🔓 Model Flexibility: Any model compatible with HuggingFace Transformers can be used—not just the ones listed above. This is a major strength of Intentioned.tech. However, you should always test models thoroughly before deploying them and check their license terms to ensure compliance with your use case and applicable regulations.
VRAM Requirements (GPU):
  • 1-3B models: 4-6GB VRAM (RTX 3060 12GB works great)
  • 7-9B models: 8-12GB VRAM (RTX 3080/3090 recommended)
  • 4-bit quantization reduces requirements by ~60%
⚠️ CPU-Only Mode: If you don't have a GPU, you'll need to use much smaller models (1B-3B parameters max) and ensure you have sufficient RAM (16GB+ recommended). CPU inference is significantly slower than GPU inference.

Configure your model in the config.json file or through the configuration tool. Models are automatically downloaded from HuggingFace on first use.

Research & Evidence

Are there any studies to back this software up?

Intentioned.tech is currently an Honors Thesis research project. While the software itself is new, it is built on well-established research in several fields:

  • Virtual Reality Social Skills Training: Studies show VR-based social practice improves outcomes for individuals with autism (Parsons & Mitchell, 2002; Kandalaft et al., 2013)
  • Conversational AI for Therapy: Research on chatbot-assisted CBT shows promise for anxiety and social phobia (Fitzpatrick et al., 2017)
  • Speech Analytics: Filler word reduction and pacing feedback have been used in public speaking training for decades (Toastmasters International research)
Disclaimer: Intentioned.tech is an educational tool, not a medical device. It is not a replacement for professional speech therapy, counseling, or clinical treatment. If you have a diagnosed condition, please consult with qualified healthcare professionals.

We are actively collecting data (with user consent) to publish formal efficacy studies. If you're interested in participating in research or collaborating academically, please contact us.

Pricing & Licensing

What is the price of the software?

Pricing has not been finalized yet. The software is currently in beta/demo phase as we gather user feedback and refine the product.

What we can share:
  • The demo on this website is free to use
  • We plan to offer multiple licensing tiers (Personal, Professional, Enterprise, Educational)
  • Academic and non-profit discounts will be available
  • Self-hosted deployment will be an option for privacy-conscious organizations

Sign up for our mailing list or follow the project to be notified when pricing is announced.

My economy has effectively dollarized! I cannot afford a GPU let alone the service Intentioned.tech provides! Can I sign up as a non-profit or similar?

We understand that many regions face severe economic challenges, including currency devaluation, hyperinflation, or de facto dollarization that makes international software services unaffordable.

📧 Contact Us Directly: If your local currency is extremely weak and standard pricing is genuinely unaffordable, please reach out to us at contact@intentioned.tech.

We may be able to offer you an affordable plan on a case-by-case basis if you can provide:

  • Valid proof of your address (utility bill, government ID, etc.)
  • Proof of your organization or business registration (if applicable)
  • Confirmation that you already have compatible hardware to run the software
Note: These special arrangements will be deprioritized compared to standard business customers regarding support response times and feature requests. However, you will still receive full access to the software and critical updates. We also do not offer these plans for EU customers due to GDPR compliance costs. Dollarization relief plans are currently not available in Ukraine, Russia, Israel, the Golan Heights, Palestine/West Bank/Gaza Strip, and Syria due to ongoing conflicts and potential economic instability that would make consistent pricing impossible in addition to potential sanctions. These regions can still get reduced pricing through academic or non-profit channels so those with no malice intent can access our software. Examples of regions where we can consider dollarization relief plans include:
  • Venezuela (Future offerings TBD)
  • Zimbabwe
  • Lebanon (Must demonstrate stable financial situation)
  • Argentina
  • Turkey (No differential pricing available, must pay in Lira converted from other currency)

Hardware & Availability

If the market gets bad for GPUs and other PC parts due to tariff issues, how will it impact me being able to obtain parts?

This is a valid concern, especially for users in the United States and Europe where tariffs on electronics can significantly impact pricing. Here's our perspective:

  • Used Market: The RTX 3060 12GB and RTX 3090 are now previous-generation cards with a healthy used market. Prices on these cards tend to be more stable. Try to avoid the RTX 20 series and earlier as these cards are very expensive and less efficient. These cards are also more gamer oriented than AI efficient as some are made before the advent of AI workloads.
  • Cloud Alternatives: For users who cannot obtain local hardware, cloud GPU services (Vast.ai, RunPod, Lambda Labs) offer hourly rentals.
  • CPU Fallback: While slower, models can run on CPU with llama.cpp for users without GPU access.
  • Lighter Models: We're optimizing for smaller models (1-3B parameters) that run on more accessible hardware like the RTX 4060.
  • Do not get a GT series GPU with a high number (e.g., GT 1030, GT 710, etc.) as these are not designed for AI workloads, extremely old by modern standards, and will perform poorly or potentially not run at all.
  • Stay Informed: Be aware of scams and price gouging in the hardware market. Many RTX 30 series cards have been replaced with GT series cards. There have been instances of sellers misrepresenting GT series cards as RTX series cards to take advantage of high demand.
  • In addition, some GPUs have been replaced with literal rocks in packaging, have had chips removed, or have been otherwise tampered with to deceive buyers.
  • If you receive a GPU that appears tampered with or not as described, report the seller immediately and avoid using the hardware. If necessary, contact the platform or payment provider to dispute the transaction. In addition, consider leaving a review or warning for other potential buyers to prevent further scams. Contact the police as well if need be. Initiate a chargeback or regional equivalent if needed to force your money back via your credit card company.
  • BUYER BEWARE: Always use CREDIT when buying GPUs for proof of purchase and seller protection. If buying from a friend, make sure you pay through PayPal or similar and enable seller protection. Pay the additional fees if needed to ensure you have recourse in case of fraud and to make sure your seller is paid the price you agreed upon. You may need a calculator for this. Google "PayPal transaction fee seller protection percentage."
⚠️ Cloud Usage Warning: If you choose to use cloud GPU services instead of local hardware, we strongly recommend selecting a GDPR-compliant provider with servers in your region. Using cloud services means your conversation data leaves your machine, which introduces privacy risks. This is inherently riskier than running locally—proceed with caution and review each provider's data handling policies.
Recommended Budget Build (2026): An RTX 3060 12GB (~€285 used) with a Ryzen 5 5600 and 16GB/32GB RAM DDR4/DDR5 provides excellent performance for Intentioned.tech.

We monitor hardware markets and will update our recommendations as conditions change.

Where do you do GPU installs as of now?

We do not currently offer hardware installation services.

Intentioned.tech is a software product. However, we understand that not everyone is comfortable installing GPUs or setting up AI environments. Here are some resources:

  • Local PC Shops: Most computer repair shops can install a GPU for a small fee ($20-50)
  • YouTube Tutorials: GPU installation is straightforward—search "how to install GPU" for step-by-step video guides
  • Pre-built Systems: Consider purchasing a pre-built PC with GPU included if you're not comfortable with hardware
Future Plans: We may offer white-glove installation services or partner with IT service providers in the future. This is not currently available.

Language Support

What languages does Intentioned.tech support?

Intentioned.tech is designed to be language-universal, especially for the EU market:

  • Speech-to-Text (STT): NVIDIA Parakeet supports 25+ languages including English, Dutch, German, French, Spanish, Italian, Polish, and more
  • Text-to-Speech (TTS): Kokoro currently supports English (US/UK). We're working on adding European language voices
  • LLM Conversations: Qwen 2.5 and Gemma models support multilingual conversation in dozens of languages
EU Language Priority: We are prioritizing Dutch, German, and French voice support for enterprise customers in the Netherlands and surrounding regions.

To configure language settings, adjust the STT language parameter in your configuration or contact us for enterprise multilingual deployments.

Legal & Privacy

Every answer here describes what the product actually does, verified against the code — not a legal conclusion about which state or federal statute's definitions it falls under. Where a compliance artifact (a signed data-processing agreement, a SOC 2 report) doesn't currently exist, this says so directly.

Is Intentioned a companion chatbot?

Intentioned.tech is a conversation-practice tool: you pick a scenario (a job interview, a parent-teacher conference, a networking event, general conversation, and others), talk to an AI playing a role in that scenario, and get feedback afterward on things like filler words, speaking pace, and how you structured what you said. Several states — New York, California, Oregon, and Washington among them — have laws that define "companion chatbot" or similar terms and impose duties (AI disclosure, crisis-referral information, break reminders, and more depending on the state) on products that meet those definitions.

This product already does several of those things regardless of the definitional question — see "Does it detect self-harm" and the AI-disclosure answer below. Whether the product meets any specific state's statutory definition is a legal determination for counsel, not something this FAQ concludes.

Is this therapy?

No. Intentioned.tech is not offered, marketed, or configured for clinical or therapeutic use. No persona in the product is intended to claim to be a licensed mental health professional — that's a standing product constraint — but as of this writing no audit has been run to confirm that none of the product's built-in or operator-configured scenario prompts actually do this. Treat this as a design intent that hasn't yet been checked, not a verified guarantee.

If you or someone you know is in crisis, the product's built-in crisis notice provides the 988 Suicide & Crisis Lifeline (call or text 988, or 988lifeline.org) — see "Does it detect self-harm" below for exactly when that fires.

Does Intentioned analyze my emotions?

No AI system in this product infers your emotional state from your voice, face, or any other signal. This was directly verified against the code:

  • The scoring dimensions that sound most emotion-adjacent — including the one that scores confident vs. hedging language — work from the words in your speech-to-text transcript only, and are explicitly instructed not to infer tone of voice.
  • The optional webcam feature ("eye contact tracking") detects whether a face is present in frame, using face-detection software. It does not read facial expressions, gaze direction, or head position — no model capable of that is loaded anywhere in the product — and raw video never leaves your browser.
  • The written feedback summary describes communication behavior you can see in a transcript (clarity, structure, directness) — it does not narrate or speculate about your personality or emotional state.

Does my voice or video leave my device?

Partial answer — a complete one is pending a full network-surface audit that hasn't happened yet.

What's verified so far:

  • Video from the optional webcam feature never leaves your browser — only a small count of "was a face detected in this frame" values is sent to the server, not video or images.
  • By default, the server this product talks to runs on your own machine or your operator's own machine, not a third-party cloud service.
  • Speech-to-text always runs locally. Text-to-speech can also run entirely locally, or can use an online service (Microsoft's Edge-TTS) — but only if online TTS has been turned on for the server. Only the AI character's own reply text is ever sent this way, never anything you said or typed. Non-English audio is not exempt from this — because the offline engine only supports English, non-English replies are more likely to use the online service once it's turned on, not less.

What this doesn't yet cover: a complete, audited list of every network call the product makes under every configuration. Treat this answer as accurate for the defaults described above, not as an exhaustive guarantee, until that audit is done.

Does it detect self-harm, and what happens if it does?

Yes, separately from and never as punishment for anything else. If something you say indicates possible self-harm, suicidal ideation, or a similar crisis, the product shows a notice — separately from anything the AI character says, labeled as coming from the app itself — with the 988 Suicide & Crisis Lifeline's number and link, plus the Crisis Text Line. This never counts against you, blocks your session, or is treated as a content violation.

Some referenced state laws additionally permit an operator to substitute a youth-specific crisis line for a user identified as under 25 — this product does not currently identify user age or make that substitution; the fixed 988/Crisis Text Line referral above is what every user currently receives, regardless of age.

Separately, if a conversation contains repeated policy-violating content (unrelated to a crisis disclosure — things like harassment or requests the AI has to refuse), the product tracks that: at 3 distinct violations in one session, a summary is saved for the operator running that server to review; at 5, the session ends. Two severe violations end a session immediately, without waiting for either count above. Because self-harm disclosures are tracked completely separately from this system, having a hard conversation about something painful does not risk hitting this threshold.

Who can see my session data?

This depends entirely on who operates the specific instance you're using — Intentioned.tech is self-hosted software, not a service the company runs and can see into by default. On the instance's default privacy track, your conversation data is retained locally for a set number of days and then deleted, and the company does not receive it. If you opt in to the alternative track (for continuity across sessions), your data is retained until you delete it or withdraw consent, and is used only for your own analytics — not shared with the company by default. Whoever operates the server you're talking to (which could be you, your employer, or your school, depending on deployment) can access the data stored on that server, same as any self-hosted software; see the questions below for what that means for a school-run deployment specifically.

Is it suitable for minors?

The product has an age-band system with an adult mode and a more restrictive teen mode, both implemented today; a further child-specific mode is designed but deliberately not turned on — specifically so it can't half-work under a label that promises protections it doesn't yet have. Which mode applies to a given user or deployment is a setting that's declared, not verified: the product does not check or estimate anyone's age, so a wrong or unset declaration isn't caught.

California and other states require a clear suitability/age disclosure for products like this regardless of a company's own assessment; that disclosure obligation exists independent of anything stated here.

What student data does Intentioned hold, if a school runs its own instance?

Whatever the school's own instance is configured to retain: by default, session transcripts and feedback are stored on the server the school runs, for a configurable number of days, then deleted. If a student's account opts in to extended history, that student's data is retained until deleted, on that same school-operated server.

The company that makes Intentioned.tech does not receive, store, or have access to this data under the default configuration — it lives entirely on infrastructure the school controls, the same as any other self-hosted application the school might run.
This is conditioned on one setting the district controls, not an unconditional guarantee: online text-to-speech. If a district turns this on, the AI character's spoken reply text — never anything the student themselves said or typed — is sent to a third party (Microsoft, via Edge-TTS) to be converted to audio. See "Does online text-to-speech affect FERPA?" below before enabling it.

Is Intentioned FERPA-compliant?

FERPA obligations run to the educational agency or institution, not to a software vendor directly — a school using Intentioned.tech in a self-hosted deployment is the one that holds the FERPA compliance obligation for the data on its own server, the same as it would for any other self-hosted tool it operates. Where a vendor relationship requires FERPA-relevant contractual terms (a "school official" designation, data-use limitations, and similar), those terms would need to be in a written agreement between the school and the company.

No such signed data-processing agreement or FERPA-specific contract template currently exists for this product. A school should not treat this FAQ, or the product's existence, as satisfying that requirement on its own.

Does online text-to-speech affect FERPA?

If it's enabled, potentially yes, and it's the district's determination to make, not this FAQ's. Under the "school official" exception districts commonly rely on for classroom software, the district must retain direct control over how a vendor and its sub-processors use and maintain education records.

Turning on online text-to-speech routes the AI character's reply text (never a student's own words) to a third party — Microsoft, via its Edge-TTS service — that the district did not necessarily contract with or account for. A vendor sub-flow the district's own FERPA analysis didn't consider is exactly what the school-official exception does not cover by default. A district should not enable this setting without first evaluating it under its own FERPA obligations.

Two things worth being precise about:

  • What's weaker than it might sound: only the AI's own generated reply text is ever sent this way. Nothing a student said, typed, or that appears in their transcript as their own words is transmitted to this third party. Speech-to-text always runs locally, regardless of this setting.
  • What's not yet built: there is currently no warning or confirmation step at the point this setting is turned on, because there is currently no control for it in the desktop configuration tool at all. The only ways to enable it today are an environment variable set before the server starts, or a direct settings-update request from any signed-in account on the server — not currently restricted to an administrator account.

Do you have SOC 2?

No. No SOC 2 report, audit, or equivalent third-party security attestation exists for this product as of this writing. No HECVAT (Higher Education Community Vendor Assessment Toolkit) response has been prepared either.

A district that requires one as part of procurement should treat this as a current gap, not something available on request.

What about New York Education Law §2-d?

A New York district evaluating this product will reasonably ask for a data privacy agreement meeting §2-d's requirements and a copy of the parents' bill of rights for data privacy and security.

Neither currently exists for this product. Both would need to be created — the parents' bill of rights is typically a document the district itself is required to publish (often with vendor-specific supplements), and the DPA is a contract that would need to be drafted and executed between the district and the company, consistent with whatever data flows the district's own instance actually has.

Does it use webcams, and can that be disabled?

Yes, optionally, and it's off by default. The webcam feature ("eye contact tracking") requires a student to explicitly click a "camera on" control in their session; it never activates automatically, and there is no separate district-level setting because it's already opt-in at the individual-session level — a student or the school can simply not enable it. When it is enabled, it detects whether a face is present in the video frame and nothing more (no expression, gaze, or identity information is computed or stored), and raw video never leaves the student's browser.

What happens if a student expresses self-harm?

The same mechanism described above for any user: a system-level notice (not from the AI character) surfaces the 988 Suicide & Crisis Lifeline and the Crisis Text Line. There is currently no age-based substitution to a youth-specific line — every student receives the same fixed referral as every other user. This is independent of the district's own crisis-response protocols and does not replace them — it is a fixed, always-on notice built into the product, not a substitute for a school's own safeguarding obligations or staff notification procedures.

A district relying on this product should have its own plan for what happens after that notice displays; the product does not itself contact school staff, parents, or emergency services.

Who is the operator of a self-hosted school instance — you, or the school?

In a self-hosted deployment, the school or district is the operator of that specific instance: they install it, run it on their own infrastructure (or infrastructure they've contracted for), control its configuration, and hold the data it generates. Intentioned.tech (the company) provides the software and periodic updates but does not host the school's instance, does not have standing access to the data on it, and is not in the data path for a properly deployed self-hosted instance running under its default privacy settings.

This is the same operator relationship a school has with any other self-hosted application, and it's the reason "who can see student data" and "are you FERPA-compliant" both come back to the school's own instance and its own agreements, not to a promise the company can make on the school's behalf.