XAI Forensics
Shows why a transformer made a decision, not just what it decided. Three-layer forensic inspection with a live demo.

I got obsessed with the gap between “the model said so” and “here’s why” and spent the last year building my way into it. IEEE-published, production-tested, and the person who asks “but can we verify that?” in every design review.
Ghaziabad, India · Open to remote / relocation
Open to Applied AI · ML Engineer · FDE roles
Shows why a transformer made a decision, not just what it decided. Three-layer forensic inspection with a live demo.
India's grid needs to know tomorrow's sun and wind together, not separately. This model does both - 2.2% error, IEEE 2025.
Every AI recommendation ships with the numbers it cites. Verifiable AI for D2C brands, with reasoning shown before you act.
Shows why a transformer made a decision, not just what it decided. Three-layer forensic inspection with a live demo.
India's grid needs to know tomorrow's sun and wind together, not separately. This model does both - 2.2% error, IEEE 2025.
Every AI recommendation ships with the numbers it cites. Verifiable AI for D2C brands, with reasoning shown before you act.
Drop any file, get a structured analysis back. Gemini 2.5 primary with OpenRouter fallback - it keeps running even when the model goes down.
Three models watching one student: one flags if they're stuck, one predicts if they'll drop out, one recommends what to study next.
Shared study rooms where everyone sees the same state. Built so the session never splits - the server owns the truth, not the browser.
Drop any file, get a structured analysis back. Gemini 2.5 primary with OpenRouter fallback - it keeps running even when the model goes down.
Three models watching one student: one flags if they're stuck, one predicts if they'll drop out, one recommends what to study next.
Shared study rooms where everyone sees the same state. Built so the session never splits - the server owns the truth, not the browser.
Multimodal sentiment analysis: novel fusion architecture combining visual and textual modalities to improve prediction on ambiguous inputs where unimodal models consistently disagree. Under research collaboration. Public writeup planned once results clear peer review.
Building a personal XAI project: a tool that explains its own reasoning as it goes, not after the fact. Separately, the Trurism , cutting irrelevant recommendations before they reach the engine.
"REALM: Retrieval-Augmented Language Model Pre-Training" (Guu et al., 2020). Thinking through how retrieval-augmented inference applies to domain-specific explainability and whether retrieved examples can ground counterfactual generation in XAI systems.
Coursework
Click for details →
Coursework
Click for details →