Department B · The Writing

Notes from the bench.

Long-form project post-mortems, an essay on the role I am currently targeting, and a few drafts left in plain sight as a form of public accountability.

Part IEditor's pick.long read · 11 min
Part IIOther essays.1 more
Part IIIProject deep-dives.27 writeups

Each project in the catalogue has a corresponding writeup: challenge, solution, impact, architecture, learnings. Below is the index. The full dossier on each one lives under /projects.

  1. 01FIFA Soccer DS.Production-shaped computer-vision pipeline for soccer video: YOLOv8n detection, ByteTrack tracking, and a GraphSAGE tactical-graph scaffold, served via FastAPI with MLflow + DVC.
  2. 02Soccer Vision Research.Modular research framework that orchestrates four pretrained vision models (RF-DETR, SAM2, SigLIP, ResNet) into one configurable soccer-analysis pipeline with selectable fusion strategies.
  3. 03WebMCP Portfolio Integration.This portfolio registers structured tools through the WebMCP browser API, so an AI agent can query it directly, search projects, read the resume, even change the page, instead of scraping the DOM.
  4. 04Nobel Data Intelligence.Physics-informed deep learning for protein stability and enzyme kinetics. A tri-modal architecture fuses ProtT5 sequence embeddings, a VDOS vibrational spectrum from normal mode analysis, and ChemBERTa chemistry through learned gated attention.
  5. 05Website Watcher.A website archiver and change-detection service: it discovers pages, snapshots them with ArchiveBox, detects content changes by hash, makes every version full-text searchable, and seals versions with Merkle trees and OpenTimestamps anchoring.
  6. 06CAG Deep Research.A causal-adversarial deep-research engine: LangGraph plans a causal graph of a question, red and blue team agents investigate each causal edge in parallel, and a dialectical judge returns verdicts. Built on a hexagonal architecture with swappable search and LLM providers.
  7. 07Biotech Accelerator.Multi-agent AI system for biotech research that routes a query through protein databases, literature, structural analysis, and chemistry, then proposes experiments worth running next.
  8. 08ContextBox.CLI-first personal knowledge assistant. Captures screenshots with OCR, pulls content from the web, embeds everything with sentence-transformers into SQLite, and answers questions over it from the terminal via GitHub Models.
  9. 09Travel Booking Platform.A 14-service travel booking platform (flights, hotels, cars, deals, billing) built as a TypeScript monorepo: independently deployable services communicating over Kafka, polyglot storage across MongoDB and MySQL, a React client with i18n, and a Python AI concierge, all orchestrated on Kubernetes.
  10. 10Kayak - Travel Metasearch.A team-built, 3-tier travel metasearch and booking platform: a React client over Node/Express services (API gateway, user, search, booking) backed by MySQL, MongoDB, and Redis, with a separate Python AI service for chat, deal scoring, and bundles.
  11. 11Airbnb - Distributed Booking.A team-built Airbnb-style booking platform that grew from a monolith into event-driven microservices: a React/Redux frontend over four Node services and a Python AI agent, talking through Kafka with MongoDB, and deployable on Kubernetes.
  12. 12Metabolic Syndrome Risk Validation.An external, longitudinal validation of two metabolic-syndrome risk scores on the US NHANES Linked Mortality File, benchmarked against established clinical equations (Framingham, ACC/AHA PCE, FINDRISC) with survey-weighted competing-risks survival analysis.
  13. 13Data Clean Env.An OpenEnv-compliant reinforcement-learning environment for tabular data cleaning. An agent fixes messy datasets through a typed action space (convert types, fill missing, dedupe, standardize, parse dates) and is scored per issue resolved. Built for the Meta OpenEnv Hackathon.
  14. 14Tailor Measurement App.A phone-first, offline-capable PWA built for a small tailor shop: record customer measurements, track multi-garment orders, manage garment templates, and queue photos offline. React + Firebase with a strictly layered, unit-tested core, in English and Gujarati.
  15. 15Rubik's Cube Timer & Visualizer.A keyboard-first speedcubing app in vanilla JS: a Three.js 3D cube, WCA-style scrambles for 2x2 to 7x7, an inspection timer with +2/DNF, full session stats (Ao5/Ao12/Ao100), and a C++ solver compiled to WebAssembly, all offline-capable as a PWA.
  16. 16CPU Scheduling Visualizer.An interactive web visualizer for 9 CPU scheduling algorithms: build a process set with CPU and I/O bursts, run any algorithm, and compare them on a Gantt chart by average completion, turnaround, waiting, and response time.
  17. 17Unified Search.A compact whitepaper plus a runnable simulator that pits five autonomous-search strategies (random walk, Levy flight, gradient ascent, surge-cast, swarm stigmergy) against each other in a noisy, sparse plume environment, with an interactive in-browser Digital Lab.
  18. 18ReceiptLens.An offline iPhone prototype that reads receipts, documents, and screenshots by running the MiniCPM-V 4.6 vision-language model fully on-device with llama.cpp, no server, no upload. A SwiftUI app wrapping an on-device multimodal engine.
  19. 19Voxt.A voice-to-text clipboard tool for Linux written in Go: press a global hotkey, speak, and the transcription (via Groq's Whisper) lands on your clipboard, with a system tray, desktop notifications, SQLite history, and Wayland and X11 support.
  20. 20Diabetes Prediction (Stacking).A stacking ensemble for early diabetes prediction on the PIMA dataset: six hyperparameter-tuned base classifiers feed a Random Forest meta-learner. Co-authored and published at IEEE AIMV 2021.
  21. 21Stock Data Platform.Unified market data warehouse, 10 tickers, 25 years of history, 4 data sources. Kafka streaming, Airflow orchestration, TimescaleDB star schema, Dockerized.
  22. 22Recognize (SAM3 Experiments).Short experiments running Meta's SAM3 (Segment Anything Model 3) on images with text prompts, plus a small script to visualize the predicted masks.
  23. 23Accurate Guesser.Java desktop guessing game (GUI).
  24. 24Basic Banking System.Simple login + transactions with PHP backend.
  25. 25Chatbot (NLP).Intent-based bot trained from JSON corpus.
  26. 26Immigration Management System.Relational schema + CRUD for immigration ops.
  27. 27Wikipedia Analysis.URL → scrape → tokenize → top frequent words with visuals.
Part IVIn the pipeline.drafts · subject to change
  • OutlinedBuilding a five-agent system in ten days, including the parts that did not work.
  • OutlinedReading vLLM end to end: notes from landing a patch in a large codebase.
  • OutlinedWebMCP, six months in: what happens when an LLM can read your site directly.
  • ThinkingVDOS spectra as a learned feature for protein stability.
  • ThinkingWhat one year of daily Cursor, Claude Code, and Codex actually changes.