Rohan Sharma

Accel AI Innovate, 4th place

An iPhone app that finds internships for UK university students, scores how well each one fits, and tracks every application from saved to offer.

• Six Claude agents scout roles and draft answers. One of them, the Guardian, checks every claim against your CV and blocks anything made up.
• Built with my team at the hackathon as an Expo app with a Supabase backend, then rebuilt in native SwiftUI for the App Store, with Pro and Max subscriptions.
• The roles feed reads employers' own career systems, such as Greenhouse, Lever and Workday, and found 110 UK spring weeks from 65 firms.
• Profiled a 5.25-second launch freeze and removed it by caching regexes and scoring about 4,500 roles off the main thread.

SwiftUI, Supabase, Claude API, Next.js

Orbit's website

XRlabs Real-Time Vision Challenge, 5th place

Finds the tip of every surgical tool in a video frame and names it as bipolar, scissors or suction, at 113 frames per second.

• Tips are points, not boxes, so the model paints a heat map. One network spots candidate tips, a second checks each one, and confident tips are pinned to the pixel.
• Trained in one day on six cloud GPUs from 1,741 images, with five videos held back as a fair test.
• The scoring made one wrong answer cost about five right ones, so I tuned for precision and lifted the score from 0.52 to 0.68.

ConvNeXt-Tiny, heat-map detection, data augmentation

AI Chessathon, 4th of 500 teams

A chess agent for an Optiver-sponsored competition where every team had to train its own model. Our team reached the semi-finals of the 40-team knockout.

• A 2,800-line alpha-beta engine compiled with Numba, plus a NumPy evaluation network trained on 4.55 million Stockfish-labelled positions.
• Every change had to win a sequential statistical test (SPRT) before it shipped. I rejected a re-tune that fitted 9.2% better but lost 117 Elo.
• Caught early-stopping bias in my own testing: a +167 Elo pass re-ran at +38 ± 44 Elo in a fixed-length match.

Python, Numba, NumPy

Code on GitHub

Personal project

A simulated exchange where four market makers compete for fills, with latency, queue priority, toxic flow, news jumps and fees.

• Tuned on 500 training days, then tested on 1,000 unseen days, where Avellaneda–Stoikov inventory skew beat symmetric quoting.
• Cut maximum drawdown to 4.4% of capital, against 37.8%, and inventory by 75%. It won in all 20 sensitivity settings.

Python, NumPy, pytest

Code on GitHub

Cambridge Battlecode, team captain, top 3 in the UK

A C++ bot for Cambridge's student AI programming competition, which has about 5,000 participants worldwide and sponsors including Jane Street, Citadel and HRT.

• I lead a two-person team. Our test harness plays about 15,000 games every 10 minutes on 384 threads.
• Ran a sponsorship drive of 600+ personalised pitches that secured $500 of compute credits.

C++

Personal project

A buy case for Melrose, the owner of GKN Aerospace, from a ten-year DCF that gives its engine aftermarket income only partial credit.

• An Excel model with 1,178 live formulas across bear, base and bull cases, plus a two-page research note.
• An independent Python version checks 376 workbook values, so the spreadsheet and the code agree.
• In the base case, all 25 combinations of WACC and terminal growth value the shares above the price.

Excel, Python

Model and note on GitHub

A learning project, not investment advice.

Personal project

Inverse-volatility weights across equities, Treasuries and gold, rebalanced monthly with trading costs and no look-ahead.

• Backtested from 2006 to 2026 against a 60/40 portfolio, after costs.
• Maximum drawdown of 15% against 27%, and alpha of 2.3% a year.

Python, pandas

CREST Gold Award

A live scoreboard on a Bluetooth LED panel and a camera tracker for the shuttlecock, used by coaches at Queensmead Sports Centre.

• Reverse-engineered the panel's Bluetooth protocol, checked against 200+ captured packets, then wrote a Python driver, Flask API and React app.
• Trained a YOLO detector on 4,000+ labelled images to 97% accuracy. Coaches used it with classes of 30.

Python, Bluetooth LE, Flask, React, YOLO

Code on GitHub

RoboMasters Hackathon, LEC Robotics, London

With a team of five, I assembled and programmed a UGOT modular robot in Python in a single evening, for a run of sensor missions that ended in an autonomous delivery.

Python, UBTech UGOT kit

Extended Project, graded A*, and TeenTech Gold

A 3D-printed knee brace with active Peltier cooling and Bluetooth control.

Embedded electronics, CAD, FEA, 3D printing

With Overbury, 1st for Best Innovation

A radial-trellis scissor lift, designed with Overbury and validated with FEA to a 250 kg static load.

CAD, FEA

Personal project

Watches public deal feeds for ASUS ROG Zephyrus laptops, new, open-box and refurbished, and alerts on big discounts.

• Scans every 30 minutes on GitHub Actions and stores listings in SQLite. When 14 of 17 posted deals turned out to have expired, I added expiry checks.

Python, SQLite, GitHub Actions

Code on GitHub