RisshabhBhardwaj

One summer each: machine learning, software and technology architecture.

1. 

   Tech Architecture Summer Analyst

   Accenture · New York City
   Jun to Aug 2026

   • Media
   • Entertainment
   • Cosmetics
   • Athletics
   • Nonprofit
   989898%

   manual reconciliation removed

   929292%+

   RAG precision, recall, F1

   What I built

   • Built and deployed a six-agent LLM system on AWS that turned a multi-week creative campaign workflow into minutes.
   • Built a schema-harmonization pipeline that cut a week of manual data reconciliation to about an hour.
   • Shipped a RAG retrieval system for a nonprofit with over 92% precision, recall and F1.
2. 

   Software Engineer Intern

   L'Oréal · Clark, New Jersey
   Jun to Aug 2025

   $4M$4M$4M/yr

   estimated annual savings

   999999%

   stability accuracy across 7,000 formulas

   What I built

   • Built AI automation pipelines saving an estimated $4M a year and making chemists 15% more efficient.
   • Replaced manual visual inspection with a computer vision and LLM pipeline that judges formula stability at 99% accuracy across 7,000 formulas.
   • Wrote an R analytics pipeline for correlation and clustering that feeds formula optimization models.
3. 

   Machine Learning Engineer Intern

   Educational Testing Service · Princeton, New Jersey
   Jun to Aug 2024

   Built TensorFlow security features for the GRE, TOEFL, SAT and AP exams, now in production in ETS Test Security as a Service.

Each project with working code has its own page. Code is linked where the repository is public. The businesses I am building are under Ventures.

• Gridiron Oracle

  Causal forecasting and execution

  A live site that forecasts every NFL player's weekly stats using only what was known before kickoff, so each forecast is one you could have made at the time.

  0.460.460.46µs

  p99 message latency

  17–2317–23×

  lower p99 than a mutex queue

  • Python
  • C++
  • XGBoost
  • LightGBM
  • CatBoost
  • FastAPI
  • PostgreSQL
  • React
  • +4 more technologies

  Breakdown: Gridiron OracleLive site for Gridiron Oracle (opens in a new tab)Code for Gridiron Oracle on GitHub (opens in a new tab)

• Memory-constrained forecasting

  Shows that a forecaster with limited memory predicts electricity-price spikes better by keeping a few learned summaries than 14 times as much raw history.

  191919%

  lower tail-forecast error at equal memory, best budget

  1414×

  more raw history beaten at the same bytes

  • Python
  • Polars
  • MLX
  • Parquet

  Breakdown: Memory-constrained forecastingRead the research report: Memory-constrained forecasting

• Prediction-market trading system

  Python signals, C++20 execution

  Spots price gaps between prediction markets and places the trades through an engine fast enough that the wait is measured in microseconds.

  10.410.410.4µs

  p50, Python to C++

  58.8k58.8kmsg/s

  single-thread peak, no dropped writes

  • Python
  • C++20
  • Cap'n Proto

  Breakdown: Prediction-market trading system

• Satellite constellations for disaster response

  Autonomous orbital optimization

  Designs satellite constellations that revisit disaster areas sooner, using a learned model in place of hours-long orbit simulations.

  43.943.943.9%

  shorter revisit gap than the standard 48-hour single-satellite benchmark

  0.8920.892R²

  on the held-out test set

  • Python
  • XGBoost
  • Gaussian processes
  • NSGA-II
  • Three.js

  Breakdown: Satellite constellations for disaster responseRead the research report: Satellite constellations for disaster responseCode for Satellite constellations for disaster response on GitHub (opens in a new tab)

• Veil

  Cross-platform federated learning SDK

  Lets an app personalize what it shows without collecting user data: the model learns on the phone or in the browser, and the raw data stays there.

  000

  raw examples leave the device

  • TypeScript
  • Swift
  • Kotlin
  • Fastify

  Breakdown: VeilCode for Veil on GitHub (opens in a new tab)

Businesses I am building or helping to build, and where each one stands today.

• Mainstay

  Live

  Gives a business one view of its calls, sales and ad spend, so the owner can see which marketing brings in customers.

  Live with a pilot venue.

  • An AI answers the phone when no one can, books and changes bookings, and hands the call to a person when asked.
  • Talk of injury, legal trouble or insurance gets no AI answer. A person is alerted right away.
  • Each business's data is walled off inside the database itself, and the app refuses to start if that wall is missing.
  • A release will not go out while a call is in progress.

  Roadmap

  1. Answers calls and books for the pilot venue: Done
  2. Serve a second and third kind of business with one shared booking and intake setup: In progress
  3. Turn on email replies and texting: Planned

  Breakdown: Mainstay

• NavisTeams

  Live

  Staffing software for multi-location businesses, on web, iOS and Android, that I am joining as a co-founder after mapping the whole system.

  Private beta, by invitation.

  • Managers build and publish schedules. Staff claim open shifts, swap shifts and ask for time off.
  • Floor staff clock in at a shared kiosk with a PIN.
  • Hours flow to QuickBooks payroll, and sales sit next to labor cost.
  • An AI assistant answers questions about hours, pay, schedules and expiring documents.

  Roadmap

  1. Scheduling, time clock, payroll sync and the AI assistant are built: Done
  2. Private beta with the first operators: In progress
  3. Switch on the AI schedule assistant, then add paid plans: Planned

• BizSi v2

  Building

  The next step of Mainstay: an AI voice agent that picks up a business's calls almost at once.

  Not taking real calls yet.

  • An owner types a phone number, website or name and talks to an agent that has already found their business.
  • Every sentence is checked before it is spoken, in English and Spanish. A sentence that fails is dropped, never rewritten.
  • What a caller says is never treated as a fact about the business.

  Roadmap

  1. Reply checks and call-safety rules are built: Done
  2. The first real call: In progress
  3. Real callers, then a pilot with businesses: Planned

• AbilityOS

  Building

  A desktop mentor that helps people learn a job by doing it: it sees the task and gives the smallest hint that helps, even offline.

  Built, not deployed. Work is paused for now.

  • Built to work fully offline, for places like correctional facilities where the internet is off limits.
  • The coach model is squeezed to about one bit per weight, so it runs on a plain CPU.
  • Hints come in steps, and a hint never gives away the protected answer.
  • It reads only the window being worked in, never password fields, and the server never receives screen images or audio.

  Roadmap

  1. Desktop app, coach, trainer and admin views are built: Done
  2. Ship a more reliable coach and polished learner, trainer and admin views: Planned
  3. A first minute of three screens or fewer: Planned

  Breakdown: AbilityOS

• Kadi

  Planning

  A marketing system that finds customers for small businesses and reports its results in booked jobs.

  Plans and decisions only. No code yet.

  • It judges itself by booked jobs and revenue, not vanity numbers.
  • It runs on my own businesses first, then beta clients, then as an agency service.
  • Nothing is sent, posted or spent until the owner approves it.
  • The goal is 30 minutes a week or less of the owner’s time.

  Roadmap

  1. Research papers and a decision book: Done
  2. Run it by hand first, for my own businesses and a few beta clients: Planned
  3. Automate what repeats, then sell it as a service: Planned

• Nani

  Planning

  Early research into a business that would serve US nursing homes.

  Research only. No product chosen yet.

  • The research ends with one idea, chosen from many.
  • Evidence comes from live sources, and real voices come first: staff, residents, families and operators.
  • No idea may make residents’ lives worse to save money.

  Roadmap

  1. Research on the problems, rules and money in nursing homes: Done
  2. Test two ideas with real buyers: In progress
  3. Pick one, and build software first: Planned

1. Through 3rd grade

   Peter Muschal School

   Bordentown, NJ

2. 3rd to 6th grade

   SciCore Academy

   New Jersey

3. 6th to 7th grade

   Bordentown Regional Middle School

   Bordentown, NJ

4. 8th grade

   John Witherspoon Middle School

   Princeton, NJ

5. High school

   Princeton High School

   Princeton, NJ

6. 

   West Lafayette, Indiana

   • BS Computer Science

     Machine Intelligence concentration

     Expected Dec 2026

   • BS Artificial Intelligence

     Minor in Mathematics and Philosophy

     Expected Dec 2026

   • MS Artificial Intelligence

     Machine Learning focus

     Expected Dec 2027

• ResumePDF
• Email[email protected]
• GitHubRisshabhBhardwaj(opens in a new tab)
• LinkedInrisshabh(opens in a new tab)