I keep building small, self-tested tools to remove friction from tasks I do myself.
Now AI agents build a lot of those tools and their interfaces, and the results often look finished while missing what the person actually needed. I want to study why: whether an agent misreads the design problem or just executes it poorly, and what design knowledge, given up front, actually helps.
Pratham Arora/AI & Automation Architect, Practus AdvisorsB.Tech. in Computer Science & Artificial Intelligence, Plaksha University '26
Before working on frontier models, my work always started with direct friction: an interface failing court-side, a manual slide bottleneck, or an elective schedule collision. Now that AI agents build much of that interface work, I want to know where their interfaces go wrong.
2024 · Ergonomics & Tools
Court-Side Basketball Live-Scoring Tool
Personal project / Campus sports
Friction Noticed
Keeping score and stats for the team on paper couldn't keep up with the pace of a game, and a wrong entry could only be crossed out, not fixed.
Tool Built
A scoring site backed by Firebase, used court-side and edited almost every week for about six months: controls moved to where they were needed during play, and an undo replaced noting down mistakes.
What Was Learned
An interface that works at a desk can fail at the pace of real use; placement and recovery from errors mattered more than features. This tool is what led to my role as Head of Technology at Athleda, the sports society, where I built the Eklavya fest site and its live scoreboard.
Iterative DesignUI Under PressureFirebase
Complete Friction Timeline Record
Court-Side Basketball Live-Scoring Tool (2024)
Friction: Keeping score and stats for the team on paper couldn't keep up with the pace of a game, and a wrong entry could only be crossed out, not fixed.
Tool Built: A scoring site backed by Firebase, used court-side and edited almost every week for about six months: controls moved to where they were needed during play, and an undo replaced noting down mistakes.
Takeaway: An interface that works at a desk can fail at the pace of real use; placement and recovery from errors mattered more than features. This tool is what led to my role as Head of Technology at Athleda, the sports society, where I built the Eklavya fest site and its live scoreboard.
AST-Fingerprinted Code RAG Pipeline (Jun 2025)
Friction: Automating code documentation across 15+ languages at 50,000+ LOC/day was driving up vector embedding costs because naive text-chunking re-embedded files even when only whitespace or comments changed.
Tool Built: An AST-based parser that computes cryptographic hashes for individual function subtrees, skipping vector re-embedding for unchanged AST nodes. Paired with a migration from FAISS to Azure Cosmos DB with IVF vector indexing for durable multi-tenant persistence.
Takeaway: Cut embedding costs by 85% in this pipeline. Hashing at the level of functions, rather than raw text chunks, made it easy to tell what had actually changed and skip the rest.
Friction: Semester-7 timetables came as PDFs split by major, so anyone taking electives from another major had to scroll across several files to find out whether their courses clashed.
Tool Built: A small site where you pick your courses and see your week, with clashes and the clashing courses highlighted - shared with my cohort as their representative. Later expanded into `plaksha-timetable`, a fuller portal with conflict detection, admin schedule uploads with change diffs, and .ics calendar exports.
Takeaway: A tool built for one semester's friction turned into something worth offering the institution: I shared the fuller portal with the Academic Office.
VR Conversational Latency Study (Late 2025)
Friction: Cloud-backed VR conversational agents leave users waiting seconds for each reply - long enough to break presence - and shaving latency off cloud LLM calls only goes so far.
Tool Built: A Unity conversational agent (Gemini 2.5 Flash + Google Cloud STT/TTS, about 2.0s average round trip) with four switchable wait states - embodied gestures, a loading spinner, verbal fillers, and no feedback - evaluated in a within-subjects study (N=18).
Takeaway: Embodied thinking gestures made the agent feel significantly more alive and human-like than a loading spinner (p < 0.05). What the agent does while you wait matters as much as how long you wait.
Athleda Player Auction Slide Generator (2026)
Friction: Preparing the annual player auction meant building a slide for every registered player by hand, matching spreadsheet rows to registration photos.
Tool Built: A browser-based slide generator (`design-auctions`) with no backend of its own: it parses the spreadsheet and the ZIP of photos in the browser, fuzzy-matches player names to photos, and exports the deck. Only project settings and themes are saved (Firestore, behind Google sign-in).
Takeaway: Doing the heavy processing in the browser meant player photos never had to be uploaded anywhere. Built after stepping down as Head of Technology, so the next organizers wouldn't face the same manual work.
VLM Physical Benchmark Capture Tool (2026)
Friction: Collecting 459 benchmark images with instrument-measured ground truth required constantly toggling between the camera app and notes app for each scale or caliper measurement, risking transcription errors.
Tool Built: An integrated iOS capture tool: on shutter press, the user selects the measurement task (e.g. weight, liquid volume, angle), enters the instrument reading in a pop-up notepad, and upon save it pairs the JPEG and TXT files locally while logging metadata directly to Google Sheets.
Takeaway: Building a small, tailored tool removed measurement friction during the initial solo data collection phase. When teammates joined with Android devices, the workflow smoothly transitioned into a collaborative division of labor (one photographing, one measuring).
Where agent-made interfaces go wrong
Coding agents now generate user interfaces from a prompt. In my own projects, an agent given only a user flow produced a complete frontend that still put the wrong things in front of the user; it did better once I described the look of the site in detail. I want to understand that gap properly.
Framing failures vs. craft failures
When an agent-made interface fails, is it because the agent misunderstood the problem - the users, their tasks, what matters most - or because it executed a reasonable plan poorly? Design research shows expert designers spend more effort scoping the problem than novices do; do agents skip that step?
What design knowledge actually transfers
Teams hand agents guidelines, design systems, example screens and expert personas. Do these fix different kinds of failure? Examples help human designers but also make them copy flaws (design fixation) - does the same happen to agents?
What agents should show while they work
In VR, a loading spinner was honest about system status but made an LLM agent feel less human than thinking gestures did. Agentic tools now work silently for tens of seconds or minutes - what should they show the person waiting?
Papers
Wait for It: A Component-Level Analysis of Latency Feedback Mechanisms for LLM Agents in VRP. Arora, A. Lodha, Siddharth S.Preprint ↗
Under review, MIT Presence (2026)
QUIVER: Benchmarking and Enhancing Physical Reasoning Abilities of Vision-Language ModelsV. Lalwani, A. Shafiq, P. Arora, M. K. Gurumurthy, P. PansariPDF ↗
AI & Automation Architect - Practus AdvisorsAutomating the growth office's reporting workflows and contributing to finance data migration for the firm's in-house ERP.
Jul 2026 – Present
Cohort Representative - CSAI'22, Plaksha UniversityCoordinated grading-component scheduling across professors; built a tool for planning the semester around elective options.
Oct 2024 – Jul 2026
Undergraduate Researcher, Vision-Language Model Benchmarking - Plaksha UniversityPrimary collector for a 459-image instrument-measured benchmark and sole author of the model error analysis.
Jan 2026 – May 2026
Head of Technology - Athleda, Plaksha Sports SocietyBuilt and maintained a sports platform serving 500+ students; mentored 2 junior developers.
Apr 2025 – Feb 2026
AI Engineering Intern - CotalityAST-hash dedup cutting embedding costs by 85%; RAG over Cosmos DB vector indexing serving 17+ enterprise repositories.
Jun 2025 – Jul 2025
Active Stack & Foundations
Distinguishing tools I actively ship with from core mathematical foundations I am currently re-deriving and strengthening from first principles.
Dedicated self-study focused on building mathematical and experimental foundations without library shortcuts:
Vectors & Vector SpacesBasic ML FoundationsCalculus (Single/Multivariable)Matrix Algebra & DecompositionsClassical ML from ScratchDeep Learning from First Principles
