Manshaan AI: Restoring Dignity to Neurodevelopmental Diagnostics

Inspiration

In 2026, we are facing a global diagnostic black hole. With over 227,000 families on open referral lists in the UK alone and U.S. waitlists stretching to 30 months, the current system is failing. Every month a child waits for a diagnosis is a lost window of neuroplasticity. We were inspired by the $461 billion annual economic burden of autism and ID. This crisis is not just a healthcare problem, but a systemic collapse of accessibility. We set out to restore dignity to the diagnostic process by replacing cold, paper-based bottlenecks with warm, empathic, and mathematically precise AI.

What it does

Manshaan is a multimodal, AI-driven diagnostic operating system for Neurodevelopmental Disorders (NDDs). It turns a 2-day clinical battery into a 20-minute digital session.

Multimodal Assessment: Patients engage in a two-way, empathic conversation (via Hume AI) and perform visuospatial tasks on a digital canvas.
Clinical-Grade Logic: Our engine uses Item Response Theory (IRT) to adapt questions in real-time, calculating a patient’s "Latent Ability" $\theta$ with significantly fewer items than static tests.
Differential Insight: The platform doesn't just "screen"; it untangles the comorbidity between Autism (ASD) and Intellectual Disability (ID), providing clinicians with a "peaks and valleys" cognitive profile.

How we built it

We built a full-stack architecture designed for clinical reliability:

Backend: A FastAPI server (Python 3.12) implementing a 3-Parameter Logistic (3PL) IRT model. The probability of a correct response is calculated as: $$P(\theta) = c + \frac{1 - c}{1 + e^{-a(\theta - b)}}$$ where $a$ is discrimination, $b$ is difficulty, and $c$ is the guessing parameter.
Frontend: A React (TypeScript) application styled with Tailwind CSS, featuring a custom drawing canvas for visuospatial testing.
AI Core: We combined Hume AI’s EVI for paralinguistic emotion detection (anxiety, distress, engagement), GPT-4o Vision for canvas analysis, and Gemini 3 Pro as the "Clinical Brain" to generate adaptive questions.
State Management: Zustand manages the session state, ensuring real-time $\theta$ updates after every interaction.

Challenges we ran into

Multimodal Latency: Syncing Hume’s WebSocket-based voice stream with our backend’s IRT calculations required a rigorous async architecture to prevent response lag.
Dependency Footprint: Deploying a heavy multimodal stack (OpenCV, SciPy, NumPy) to serverless environments like Vercel hit strict 250MB memory caps, forcing us to optimize our build and pivot toward Railway.app.
Clinical Accuracy: Fine-tuning the LLM to generate valid clinical questions that map specifically to 5 cognitive domains (Memory, Executive Function, etc.) required prompt engineering that respected "Non-Device CDS" regulatory guardrails.

Accomplishments that we're proud of

End-to-End Multimodal Fusion: We successfully built a system where a user's voice tone and visual input directly affect the math of their cognitive score.
IRT Validation: Our scoring engine successfully converges on an ability estimate ($\theta$) within just 10-15 items, proving the massive efficiency gain over static paper tests.
Solving a Problem Close to Home: We've successfully designed a test, that if administered to Jasjeev's younger brother 10 years ago, would have saved his family and his doctors upwards of $50k.

What we learned

We learned that "empathy is a data point." In neurodevelopmental care, a child's anxiety is just as diagnostic as their test score. We also gained deep experience in psychometrics. Finally, we learned the importance of independent review in AI; building "explainable" dashboards is key to gaining clinician trust.

What's next for Manshaan AI

Clinical Correlation Study: Validating our IRT scores against the ADOS-2 gold standard in a formal pilot.
CPT Reimbursement Integration: Building out the billing engine for CPT Code 96146, allowing clinics to use Manshaan as a profit-center.
Longitudinal Tracking: Moving from "one-time" diagnosis to monthly monitoring, creating the world's first longitudinal dataset of neurodevelopmental progress.