Hitesh Pranav

Problem

Finding the right research advisor means reading through dozens of faculty pages. A plain keyword search misses faculty who describe their work in different words, and a black-box LLM answer can't be checked against the underlying data.

Approach

A hybrid retrieval pipeline combining BGE-Large dense embeddings (1024-dimensional) with BM25 sparse keyword matching, so results capture semantic meaning and exact term matches together. A local Qwen-2.5-0.5B model generates a short explanation for each match. Flask backend, React frontend, with client-side PDF parsing so a resume can be matched directly.

Engineering detail

Embeddings, retrieval and explanation generation all run locally instead of calling an external LLM API per query — the system stays self-contained with no per-query dependency on a hosted model.

Outcome

A working search system across faculty on PES University's RR and EC campuses that returns ranked, explained matches instead of a flat keyword list.

Problem

CAN bus, the protocol most vehicle components use to talk to each other, has no built-in authentication — any device on the bus can inject messages, which makes it a real attack surface for connected vehicles.

Approach

An intrusion detection pipeline built around a Random Forest classifier trained to flag anomalous CAN traffic, paired with a real-time Python firewall that can act on detections as they happen. python-can simulates bus traffic and attack scenarios for testing without physical hardware.

Engineering detail

The interesting part isn't the classifier — it's closing the loop: detection feeds directly into a firewall that intervenes on live traffic, rather than just flagging anomalies after the fact.

Outcome

A working detection-and-response pipeline, later extended during a research internship at C-ISFCR into a fuller cybersecurity framework for automotive networks.

Problem

A race team needs to see what's happening to the vehicle as it happens — acceleration, position, speed — not reconstruct it afterward from logs.

Approach

Built during Ignition 1.0, an 18-hour hackathon run by Team Vegavath PESU and sponsored by Ather Energy. The dashboard streamed live IMU acceleration data (Ax, Ay, Az), GPS-based position tracking and a 720p MJPEG video feed, with speed estimated by fusing IMU and GPS instead of trusting either sensor alone. Driving behavior was inferred from acceleration-variance analysis. Frontend built with Chart.js and Leaflet.

Engineering detail

The hardware had to stay light enough not to affect the vehicle it was measuring — it ended up the second-lightest system at the event while still handling live sensor fusion and video.

Outcome

5th place overall among competing teams, and a working answer to a real constraint: instrumentation that measures a vehicle without meaningfully changing its weight.

Problem

JoEL needed a real digital presence for its flagship events and student projects, not a placeholder page.

Approach

Built with Next.js, TypeScript and Tailwind CSS, structured around the lab's actual events — HackeZee and Roadshow — and its student project showcase. Deployed on Vercel with maintainability treated as a first-class requirement, since a different set of students maintains it after handoff.

Engineering detail

It isn't a personal or academic project — it's production software a club runs, edited by people who didn't build it.

Outcome

Live and in active use as JoEL's official site.