I occasionally write about Artificial Intelligence, backend engineering, and my experiences building autonomous agents.
During the Odoo Hackathon 2026, our team set out to engineer an enterprise-grade logistics and transport operations platform in just 8 hours. While many hackathon projects cut corners on data integrity, we decided to tackle one of the most notorious problems in logistics engineering: race conditions in vehicle and driver dispatching.
In real-world transport operations, concurrent requests can easily lead to double-booking—assigning a vehicle that is already en route or scheduling a driver whose commercial license has expired. To solve this decisively, we architected TransitOps around a strict 3-tier foundation utilizing React (Vite), Express REST APIs, and PostgreSQL 18 managed via Prisma ORM.
Rather than relying on UI-level validation or eventually consistent cloud databases, we embedded our business rules directly inside ACID database transactions (`prisma.$transaction`). Every state change follows a strict lifecycle (`DRAFT → DISPATCHED → COMPLETED`), atomically updating vehicle status (`AVAILABLE ⇄ ON_TRIP`) and checking driver compliance constraints in a single serializable operation.
Equally important was designing a clean 4-tier Role-Based Access Control (RBAC) matrix. Fleet Managers, Safety Compliance Officers, Drivers, and Financial Analysts each operate in tailored workspaces with domain-specific guardrails—ensuring drivers never see sensitive financial overheads while safety officers retain sole authority over driver license compliance.
Building high-integrity backend systems under extreme time pressure reaffirmed a core engineering truth: clean data modeling and transactional guarantees are not "nice-to-haves"—they are the bedrock of reliable software.
As a Data Scientist and Backend Engineer, I wanted a place where I could showcase my projects, skills, and learning journey beyond a traditional resume.
That's why I created this portfolio. The goal wasn't just to build a website—it was to create a digital space where I could document my work in Artificial Intelligence, Retrieval-Augmented Generation (RAG), Natural Language Processing, and scalable Backend Engineering.
Through this portfolio, I plan to share:
This is just the beginning, and I'll continue updating the site as I learn, build, and grow.
Thanks for visiting and being part of the journey.
I am currently drafting a research paper that introduces a novel composite evaluation metric for analyzing internal reasoning in transformer language models.
Modern language models achieve incredibly high scores on standard benchmarks, but existing evaluation metrics predominantly reward surface-level text fluency. Because of this, models often learn statistical shortcuts rather than learning to actually reason—an issue widely recognized as the "Clever Hans" effect.
To address this, my upcoming paper proposes a framework that shifts the focus away from just looking at a model's generated text. Instead, we mathematically probe the internal representations to evaluate things like how effectively the model disentangles word senses, whether its internal mechanisms actually align with logical dependencies, and how stable those representations remain under adversarial conditions.
One of our most interesting preliminary findings reveals a stark dissociation: models can possess surprisingly robust internal representations for certain semantic tasks, yet completely fail to utilize those same representations when forced to apply logical reasoning.
The paper is currently being finalized for double-blind peer review. I am keeping the exact algorithms, metric formulations, and experimental data under wraps for now to protect the integrity of the submission process, but I can't wait to share the full methodology once it's published!