Assessment-First Mental Health AI

Your instinct that off-the-shelf models give no differentiated value is half right. A raw model with a system prompt is worth nothing — it is a commodity anyone clones in a weekend. But the missing ingredient is not weights. It is structure: a validated assessment protocol, a curated and citable resource corpus, a crisis-detection layer that works, and a referral network of real clinicians. All four are product work, not training work.

That last row is the whole argument. Model inference is a rounding error against the price of the service you are selling. Spending six figures on training to shave a cost that is already 2–6% of revenue is the classic first-time-founder trap.

You described this loosely. Here is the tightened version you can put in front of a clinician, an investor, or an engineer without it being picked apart.

The problem

India has roughly 0.75 psychiatrists per 100,000 people^88,90 and a treatment gap above 83% for common mental disorders.^88,89 Kerala is the best-served state in the country at ~1.2 per lakh⁸⁸, and it is still nowhere near enough. People who need help do not know what they have, what kind of help matches it, or who near them is qualified to give it. The bottleneck is not only therapy hours — it is triage.

The product

A conversational system that conducts a structured, evidence-based psychological intake in the user's own language; produces a defensible formulation of what the person is experiencing and how severe it is; and then routes them — to self-directed evidence-based material, to a licensed human clinician, or to a crisis service — with the reasoning shown.

What it explicitly is not, in v1

• It does not diagnose. It screens and formulates, then says so.
• It does not deliver therapy autonomously. That framing is already illegal in three US states^65,67 and legally murky in India.⁶⁰
• It does not replace the clinician. It fills the 167 hours a week the clinician isn't there⁹⁴, and gets the right people into the clinician's 1 hour faster.

Why the fine-tuning instinct is wrong here specifically

The comparative literature is unusually consistent: retrieval beats unsupervised fine-tuning for knowledge injection, both for knowledge the model saw in training and for genuinely new knowledge^21,22, because retrieval supplies context relevant to the question rather than diffusing facts into weights. Benchmarks of commercial fine-tuning APIs find the same thing — they are poor at infusing new knowledge^19,24,25. Fine-tuning also remains sensitive to prompt variation and prone to overfitting.^21,23,26

There is a second, product-specific reason. In a mental health product the answer "here is what the WHO mhGAP guide says about moderate depression, and here is the passage" is worth more than a fluent unattributed paragraph. A fine-tuned model structurally cannot do that. RAG does it by construction.

What fine-tuning is genuinely good for — later

• Voice and therapeutic style. Reflective listening, Motivational Interviewing phrasing, not jumping to advice. This is behaviour, which is exactly what SFT teaches well.
• Malayalam register. Getting a model to sound like a Malayali counsellor rather than a translated American one.
• Small models, narrow jobs. Comparative work on lightweight LLMs for counselling tasks shows small tuned models are competitive on bounded sub-tasks and not on open conversation³⁶.
• Cheap specialised classifiers. A small fine-tuned model for risk detection or symptom tagging is a great use of QLoRA — and costs a fraction of a general-purpose fine-tune.

Existing work proves the pattern is feasible: MentaLLaMA fine-tuned LLaMA-2 on a 105K-sample instruction dataset (IMHI) and its 13B chat variant matched or beat state-of-the-art discriminative methods on 7 of 10 held-out mental-health test sets.^27,28,29 Note what that model does though — it analyses social media text for mental health signals. It is a classifier lineage, not a counsellor. That is the realistic target for your own fine-tune, and it is a v2 target.

Numbers below use E2E Networks' published on-demand H100 rate of ₹362/GPU-hour and their committed rate of ₹155.90/GPU-hour at 3+ months.² Assumptions are stated so you can re-run them.

Route B1 — pre-training your own model from scratch

Napkin math, using the standard 6 × params × tokens FLOP estimate. Target a Sarvam-1-class 2B-parameter Indic model⁷² trained on 4 trillion tokens: 4.8 × 10²² FLOPs. At a realistic 40% MFU on H100 (~160 effective TFLOPS) that is roughly 83,000 GPU-hours.

Route B2 — LoRA / QLoRA on an open 8B model

This is the affordable version, and the numbers are genuinely small. A typical LoRA supervised fine-tuning run on a 7–8B model lands around 100 GPU-hours.^12,13 QLoRA at 4-bit needs roughly 4GB for weights on an 8B model, so it fits comfortably on an L40S or even an L4.^10,12

The GPU bill is not the expensive part. It is roughly ₹5–8 lakh, consistent with published fine-tuning budget guides¹¹. The expensive part is the dataset — see §07 — and that is where fine-tuning projects actually die.

Pulled from E2E's own pricing¹ and product^2,3,4 pages in September 2026. Their general pricing table renders in USD while the H100 product page renders in INR; the two are broadly consistent (H100: $3.77/hr ≈ ₹335 vs ₹362 listed). Confirm every figure with their sales team before you budget against it — cloud GPU pricing in India moved significantly through 2025–26^5,6,7,9 and marketing pages lag.

* INR figure from E2E's H100 product page, which also advertises ₹155.90/hr on a 3+ month commitment. Third-party comparison sites quote E2E H100 spot as low as ₹70/hr and H200 spot at ₹88/hr^6,7,8 — treat spot claims as unverified and unsuitable for anything that must not be pre-empted. Monthly rates are already discounted below 730 × hourly.

What to actually rent, and when

• Nothing, for v1. You are calling an API. Your GPU bill is zero.
• One L4 or L40S when you want to self-host a small guardrail classifier (Llama Guard-class) or a Malayalam embedding model. ₹51–107/hr, and you can run it only during business hours.
• H100 by the day, on-demand, when you run fine-tuning experiments. Never commit to 3 months until you have run at least one successful experiment on-demand.
• The real reason to use an Indian provider is not price — it is data residency under the DPDP Act.^57,58 That is a good reason, and it applies to your database long before it applies to your GPUs.

Modelling assumption: one counselling session ≈ 30 turns, ~4,000 input tokens per turn (system prompt + retrieved passages + history) and ~300 output tokens per turn. That is ~120K input and ~9K output tokens per session. Prompt caching cuts the input side dramatically on a stable prefix; figures below are the uncached worst case. Per-token rates are Anthropic's published list prices.²⁰

Three levers that matter far more than model choice:

• Prompt caching. Your system prompt, assessment protocol and retrieved corpus chunks are a stable prefix. Cache them and the input side of that table drops by roughly an order of magnitude on cache hits.
• Model tiering. Run the conversation on a mid-tier model, but route risk classification and the final clinical formulation to the strongest model you have. Most turns don't need the expensive one.
• Batch API at 50% cost²⁰ for anything non-interactive — nightly summarisation, eval runs, corpus processing.

You called data aggregation "hectic". It is worse than that. It is the single hardest, most expensive and most legally exposed part of the custom-training path, which is the strongest argument for not taking that path yet.

What public datasets exist, and what they're actually good for

What building your own dataset costs

• Domain-expert clinical annotators run $50–150+/hour internationally; published mental health research budgets around $45/hour for psychology experts versus $25–30 for general research assistants.^37,38 In India, budget ₹1,500–4,000/hour for an RCI-registered clinical psychologist's time.
• Expert annotators cost 5–15× general crowdworkers, and mental health datasets typically need dual verification — a clinical psychologist and a second reviewer signing off.^37,38
• A defensible 10,000-sample instruction set with dual clinical review is a multi-crore, multi-year undertaking. This is why MentaLLaMA is built from social media rather than clinic transcripts.

What you should build instead, now

Not a training set. A gold evaluation set, which is 50× smaller and 100× more valuable at this stage:

• 300–500 realistic Kerala-context user scenarios, written by your clinical advisors, spanning presentation type, severity, age, language and risk level.
• For each: the correct triage outcome, the red flags that must be caught, and the responses that would be unacceptable.
• A dedicated adversarial safety subset — indirect suicidal ideation, medication questions, disclosure of abuse, psychosis, minors. This is the set that catches the failure mode described in §09.

This costs perhaps ₹3–6 lakh of clinician time and becomes the asset everything else is measured against. It is also the artefact that makes you credible to a regulator, a hospital partner and an investor.

Legally usable clinical content for your corpus

Every user turn passes through this pipeline. The safety layer runs outside the model, on both sides, because a model that is asked to police itself will fail exactly when it matters.

A licence check before you self-host anything

If you ever put an open-weight model in the serving path, read its licence as carefully as you read the clinical guidance. Llama's community licence requires a separate grant above 700M MAU, mandates a "Built with Llama" notice and a model name beginning with "Llama" on anything you redistribute, forbids using its outputs to improve non-Llama models, and carries an acceptable-use policy whose own technical reports name unqualified medical advice as a listed risk⁶⁹. Gemma has no user cap but grants Google the right to restrict usage remotely for prohibited-use violations. Qwen3 ships under Apache 2.0, which is the cleanest of the three; older Qwen releases under the Tongyi Qianwen licence cap commercial use at 100M MAU and require a separate licence for model-as-a-service. Acceptable-use policies are an under-read constraint on exactly this product category⁶⁸.

Component choices

The consequences are no longer hypothetical. Google and Character.AI agreed in January 2026 to settle multiple lawsuits from families of teenagers who died by suicide or self-harmed after interacting with its chatbots — among the first settlements alleging AI contributed to teen suicides.^49,50 Similar litigation against OpenAI is ongoing. Character.AI has since barred under-18s from open-ended chat entirely.

Non-negotiables for v1

1. Input classification before generation. Every user message is screened for risk before the conversational model is invoked. Indirect ideation — the bridges case — is the specific thing you test for, because direct ideation is easy and everyone catches it.
2. A hard-coded crisis path. On a positive risk signal the conversation leaves the LLM entirely and enters a scripted flow: acknowledge, stay present, surface Tele-MANAS 14416⁹⁵ and Kiran 1800-599-0019, offer to connect to a human. The model does not improvise here.
3. Age gating. Minors are a different legal and clinical regime. If you are not building for them deliberately, exclude them explicitly.
4. Scope refusal. No medication advice, no dosage, no diagnosis language, no drug interactions. Enforced at Gate 2, not by prompt.
5. Human-in-the-loop review. A clinician reviews a sampled fraction of sessions and 100% of escalations, weekly. This is also your training data flywheel.
6. An adversarial red-team suite run in CI on every prompt or model change. Automated clinical red-teaming frameworks for mental health LLMs now exist in the literature⁵⁴ — use one rather than inventing your own taxonomy.
7. Incident logging and reporting. MeitY's AI Governance Guidelines explicitly favour an incident-reporting posture.⁶² Build the log from day one; retrofitting it after an incident is how companies end up with nothing to show.

India

Export markets — matters now, not later

Illinois enacted the Wellness and Oversight for Psychological Resources Act on 1 August 2025: AI may not provide therapy or make therapeutic decisions, and advertising a product as "AI therapy", "chatbot counselor" or "virtual psychotherapist" is unlawful without direct licensed-clinician oversight. Civil penalties to $10,000.^65,66 Nevada and Utah have comparable restrictions.⁶⁷ The law does not ban mental-health apps — it bans the claim.

Given your reference points are Kerala platforms, assume Malayalam is a first-class requirement, not a translation layer. Three findings matter.

Language bias is a documented clinical harm, not a UX gap

Research on LLM chatbots for mental health support found users were compelled to switch to English despite it not being the language they were comfortable expressing distress in — directly hindering their ability to describe what was happening to them.⁷⁷ The same work found models' advice repeatedly failed to understand how familial dynamics differ from Western contexts, producing recommendations that would not work in users' actual lives. A systematic review notes few studies discuss cultural adaptation strategies at all, and none address cross-cultural generalisability.^35,78

Indic model options

Your three references map onto three different positions, and one of them is close enough to your stated plan that it changes the strategy conversation.

Market context: India's mental health market was ~US$2.81 billion in 2025, projected to US$6.59 billion by 2033 at ~11.3% CAGR.⁹¹ Indian trials also show lay counsellors and community health workers, supported by simple digital tools, deliver effective depression treatment in routine primary care⁸⁹ — which is a strong hint about where an assessment tool creates the most leverage.

It can work

The Dartmouth Therabot trial (NEJM AI, March 2025) is the first RCT of a generative-AI therapy chatbot: 210 adults with clinically significant major depression, generalised anxiety, or high eating-disorder risk, randomised to 4 weeks of Therabot or a waitlist. Results: 51% reduction in depression symptoms, 31% in anxiety, 19% in eating-disorder concerns, with high engagement and therapeutic-alliance ratings comparable to a human professional. The effect sizes are in the range of RCTs of 16 hours of human-delivered psychotherapy, achieved in about half the time.^79,80,81

Note who built it: a Dartmouth team including a board-certified psychiatrist and a clinical psychologist, over years. The clinical team is not an accessory to the engineering — it is the moat.

It routinely doesn't

• Stanford FAccT 2025 — stigma and the bridges failure (§09).
• Broad evaluations across mental illnesses find performance varies sharply by condition and task, so a single headline accuracy number tells you almost nothing about the cases that matter³⁴.
• Woebot shut down its consumer chatbot on 30 June 2025 after serving 1.5M+ people. The founder's stated reasons: the cost of FDA regulation, and the impossibility of integrating LLMs into a regulatory framework that doesn't exist.^82,83 The most scientifically rigorous player in the category exited for regulatory reasons, not technical ones.
• Character.AI / Google settlements, January 2026 (§09).
• Meanwhile Slingshot AI launched "Ash" in July 2025 on $93M from a16z and others⁹⁸ — capital is flowing in even as the careful operators leave.

The assessment path specifically has support

This is the encouraging part for your actual plan. Structured prompting raised diagnostic accuracy from 89.8% to 91.7% and triage accuracy from 76.8% to 86.2% in a clinical vignette study.^84,85 Multi-agent systems that process interview transcripts with separate qualitative, quantitative, judge and meta-review agents are producing credible severity estimates against PHQ-8.⁸⁷ Grounding detection in PHQ-9, BDI-II and DSM-5 symptom targets improves interpretability.⁸⁶ Every one of those results supports the structured-assessment architecture in §08 — and none of them required a custom-trained foundation model.

1. What is the claim? "Assessment and referral under clinical oversight" or "AI counselling"? This single choice sets your CDSCO exposure, your insurability, your US optionality and your architecture.
2. Who is the named clinical lead? Not an advisor on a slide — a psychiatrist whose registration is on the product and who reviews escalations. Without this the safety layer has no owner and the Illinois-style "clinical oversight" carve-out does not apply to you.
3. B2C, B2B, or referral? Campus and corporate wellness (Oppam's channel) has a buyer with a budget and a much lower per-user safety surface than open consumer signup.
4. Minors: in or out? Koott serves 13–17. If you do too, you inherit the exact risk profile that produced the Character.AI settlements. Decide deliberately, in writing.
5. Malayalam-first or English-first? Malayalam-first is the real differentiator against Wysa and the frontier apps, and it is also where retrieval quality is weakest. It is the harder and better choice — budget for it explicitly rather than treating it as a localisation task.