Anurag Bheemappa Gnanamurthy

Multi-agent job search assistant with human approval gates

• Job search tools that write for you risk fabricating the application, and the ones that do not leave the search itself entirely manual.
• A multi-agent pipeline on Google's Agent Development Kit. An orchestrator routes between Discovery, Resume Parser, Eligibility and Draft Coach, with the resume-to-scoring handoff hardwired as a sequential step so it cannot be re-routed by the model. Safety guardrails are callbacks attached to the agents rather than prompt instructions.
• A Guardian gate pauses the runner before any external send, so nothing is transmitted without explicit human approval. The resume parser extracts a PII-free profile locally, so raw resume text never reaches the model. A four-case eval suite scores tool trajectory alongside an LLM judge, covering the routing and refusal behaviour.

Full-stack learning platform with role-based access

• A learning platform's three roles see the same courses differently, and the access rules are the hard part rather than the CRUD underneath.
• Role-based access for students, faculty and administrators over a REST API. CRUD endpoints in Node and Express with Mongoose schemas against MongoDB Atlas. The API is a separate service, deployed apart from the frontend.
• Deployed and reachable, with the frontend and the API running as separate services.

ATS resume analyzer with skill gap scoring, in progress

• An applicant cannot see what an applicant tracking system sees. The score, and the skills a job description asks for that a resume never mentions, are both invisible at the point where they decide the outcome.
• A three-stage pipeline: parsing pulls structure out of a resume PDF and a job description, scoring embeds every bullet with all-MiniLM-L6-v2 and benchmarks it against a TF-IDF cosine baseline to produce a match score and a skill gap, and generation assembles a tailored PDF with the missing skills injected. A FastAPI service exposes analyse and download endpoints, and a Next.js single-page frontend runs upload to results to download.
• Runs end to end locally: a resume PDF and a job description in, a match score with skill gap analysis and a tailored PDF out. Not deployed yet.

Gaze-driven robot control on commodity webcam hardware

• Robotic interfaces assume hand or voice control, which excludes the people most likely to need an assistive robot. Dedicated eye trackers cost more than the robot itself.
• Facial-landmark detection on a live webcam feed classifies pupil direction as left, right or forward, using OpenCV and dlib. A serial integration layer carries those classifications from Python to an Arduino Uno over PySerial, which drives an L293D motor driver and steers the robot.
• Steers the robot from a commodity webcam with no dedicated eye tracker, as a working end-to-end prototype, never deployed.