I am

PRATHAM
SHAH

Dev += research · INDIA

I build lean web products and stay one arXiv binge ahead of the curve.

(01) About

I like building things that feel fast and considered: lean web products that range from static sites to AI agents, made with real care so the code won't embarrass me later. The rest of my time goes to papers I don't strictly need to read, mostly because I can't help it.

(02) Experience

Blink Analytics

LLM Intern

May 2025 to July 2026

  • Built a phone-call AI interview agent: which can serve as a reminder agent and screening agent.
  • Built bulk candidate processing: recruiters can upload and process large sets of prospects instead of handling them one by one.
  • Built payments and interview credits: companies buy plans/credits and those credits control how many AI interviews they can run.
  • Shipped ATS and recruiter tools: resume scoring, candidate profiles, job open/close, search/filters, candidate emails, and interview reports (including PDF).
  • Improved how quickly interview reports are ready after a session, and cleaned up backend Docker deploys.
  • Improved the existing live interview room: better speech recognition, AI agent model routing, noise cancellation, session reliability, and network monitoring.
  • Reviewed and helped merge teammate pull requests across company repos.
  • Optimized language models using Reinforcement Learning from Human Feedback (RLHF) and Supervised Fine-Tuning (SFT).
  • Integrated optimized LLM pipelines into internal recruitment tools with cross-functional teams.

Impact 62 merged PRs · 470+ commits across company repos

(03) Selected Work
  • A001

    Pikd

    NEXT.JS · FACE RECOGNITION
    Live ↗ Code ↗

    Event photo delivery with face recognition: guests scan once and collect every photo they appear in, in seconds.

  • A002

    Recursive Reasoning (Small Language Model)

    SLM · RECURSIVE TRANSFORMER
    Live ↗ Code ↗

    Recursive Transformer SLM for Sudoku that re-processes its own output in shared-weight thinking loops, based on “Less is More: Recursive Reasoning with Tiny Networks,” with a Streamlit dashboard for loop-by-loop visualization.

  • A003

    Reflect Reads

    MERN · BOOKSTORE
    Code ↗

    Full-stack book browsing and ordering: React/Vite + Tailwind frontend, Express REST API, MongoDB, and JWT auth.

(04) Research
  • R001

    Violent HAR Detection

    CV · YOLO-POSE · LSTM
    GitHub ↗

    Fusion model for human activity recognition: YOLO-Pose + LSTM for temporal pose features and RGB frames + MLP for spatial features to detect violent actions.

  • R002

    Hazard Surveillance and Maintenance Car

    IOT · CV · ESP-CAM

    IoT and computer vision system on an ESP-CAM-based robotic car to detect five classes of electrical hazards for maintenance and surveillance.

Publications

  • AI Based Prediction for Panic Disorder Detection Using Social and Clinical Factors

    IEEE ↗

  • Smart Detection of Passive Smokers using Machine Learning

    IEEE ↗
(05) Education

BITS Pilani, Pilani Campus

ME Computer Science

In progress

2026 to 2028

Pandit Deendayal Energy University (PDEU)

B.Tech Computer Engineering

2022 to 2026

CGPA 9.562

Selected coursework

  • Data Structures and Algorithms
  • Operating Systems
  • Database Management Systems
  • Computer Networks
  • Machine Learning
  • Artificial Intelligence
  • Computer Vision
  • Big Data
  • Information Retrieval
  • Software Engineering
(06) Open source & writing

  • Run LLM judge when --evaluate is passed

    PR ↗

    Tracer-Cloud / opensre

    Merged fix for OpenSRE: wired the --evaluate delivery path to call run_opensre_llm_judge() when a rubric is present, instead of skipping evaluation silently. Added regression tests covering the happy path, skip guards when no rubric exists, and error fallback behavior. Closes #2070.

  • Attending to all attentions (a try!) (Part 1)

    Article ↗

    Medium

    Part 1 of a 3-part attention series tracing the history from seq2seq bottlenecks to modern variants. Explains Bahdanau (additive) and Luong (multiplicative) attention, scaled dot-product self-attention, multi-head and cross-attention, and the O(n²) quadratic bottleneck—with intuition, math, and PyTorch snippets for each mechanism.

(07) Side repos

Built for me first. Grab anything that saves you time too.

  • Vanilla Transformer

    Code ↗

    PyTorch “Attention Is All You Need” from scratch for English-French translation.

  • Frame Tracker

    Live ↗ Code ↗

    Streamlit app to upload videos, scrub frame-by-frame, jump quickly, and export frames as PNG.

  • CartPole

    Code ↗

    Deep Q-Network agent in PyTorch trained to solve OpenAI Gym’s CartPole-v1.

  • Snap Log

    Code ↗

    Browser console logger (Chrome extension) that captures serialized objects, network errors, and stack traces at full fidelity.

  • pluck

    Code ↗

    Client-side extractor (Chrome extension) for images, GIFs, and video embedded in DOCX, PPTX, and PDF. No upload; runs in the browser.

(08) Shelf

Articles, videos, repos, and papers I recommend: a running list, not a ranked catalog.

Open shelf →