Aerospace Engineering · University of Bath · Year 2

I turn engineering problems into things that ship.

Second-year aerospace engineer at Bath. First author on a battery machine-learning paper, co-founder and CTO of a live delivery platform, aerodynamics lead on a competition UAV. Twenty shipped, and counting.

Looking for a 2027/28 placement year in aerospace, defence or wherever hard engineering meets software.

  1. First-author paper
  2. Winkit CTO
  3. UAV aero lead
  4. Fifth9 red team
  5. Drive-shaft design
  6. 20+ shipped
  • 0.57%blind-test SOC error, first-author paper
  • CTOco-founder, Winkit, live at winkit.uk
  • 30%lift-to-drag gain, Team Bath Drones
  • ~30PRs merged to production at Fifth9
  • FirstYear 1, 73.8% average
  • Python
  • PyTorch
  • NumPy
  • three.js
  • Autodesk Inventor
  • MATLAB
  • Arduino
  • Stripe
  • Playwright
  • Apps Script
  • Vercel

01 / Selected work

The work I would lead with.

Five builds that carry the most weight. The rest live further down.

Paper in preparation · target: Energy and AI

Battery ML research · first author

Reading a battery's charge from sound.

Coulomb counting drifts and the voltage curve sits flat. So I taught an LSTM to read state of charge from ultrasound, tested blind on real drive cycles.

  • 0.57% mean absolute error on a blind real-world Paris electric-bus cycle.
  • 42% less error on US06 once ultrasound joins voltage, current and temperature.
  • A fully blind protocol: one real vehicle recording held out of every step, model selection included.
Details
  • US06: 1.36% down to 0.79%, better on all ten of ten seeds (p = 0.002).
  • Blind Paris: 0.70% down to 0.57% for the same fusion, a 19% cut.
  • Built from first principles: a NumPy network with hand-derived backpropagation and Adam, then PyTorch LSTM ablations over features and architectures.
  • Python
  • NumPy
  • PyTorch
  • LSTM
  • feature ablation
  • 10-seed blind test

First author · supervised by Dr Christopher Vagg and Rob Gray, Institute for Advanced Automotive Propulsion Systems

Code private until publication

fig.2 · feature ablation, blind test
LSTM SOC(t−1) V, I, T + ToF, swelling (t−1)LSTM SOC(t) V, I, T + ToF, swelling (t)LSTM SOC(t+1) V, I, T + ToF, swelling (t+1)h, ch, c
The model: an LSTM unrolled over the drive cycle. Memory carries from step to step, because state of charge is an integral.

Tap a bar for its mean, spread and seed count.

Blind-test mean absolute error (percent, mean and standard deviation over ten seeds) and the share of seeds that beat the baseline, by feature variant
VariantParis MAEParis seeds improvedUS06 MAEUS06 seeds improved
VIT (baseline)0.70 ± 0.16n/a1.36 ± 0.13n/a
VIT + ToF0.69 ± 0.3060%1.36 ± 0.1040%
VIT + velocity1.02 ± 0.5030%0.91 ± 0.15100%
VIT + displacement0.67 ± 0.1870%0.84 ± 0.09100%
VIT + ToF + displacement0.57 ± 0.1980%0.79 ± 0.09100%
VIT + all ultrasonic + displacement0.75 ± 0.2360%0.84 ± 0.08100%
VIT + shuffled velocity (null control)0.70 ± 0.2170%1.32 ± 0.1670%
Blind-test mean absolute error, ten seeds, deployed estimate. The null control shuffles the velocity channel in time: if it scored as well as the real thing, the model would be learning noise. It does not.
Live · winkit.uk

Winkit · local food delivery · co-founder and CTO

A delivery platform, co-founded and built end to end.

Aggregators take 38 to 40% of a local kitchen's revenue. So three of us built our own, and I built and run the whole stack.

  • £0 commission, 0 drivers to hire, 0 new hardware for the kitchen.
  • Orders land on the kitchen's own system, live on two restaurant platforms.
  • Live Stripe payments and a GPS driver app with shifts and real-time order state.
Details
  • Card payments gated to each kitchen's real opening hours.
  • Run lean on free tiers: gating a 24/7 database cron by opening hours cut compute burn by about 68%.
  • Neon Postgres, Vercel functions and per-kitchen order adapters.
  • full-stack
  • Stripe
  • Vercel functions
  • Postgres
  • GPS dispatch
  • order adapters

Co-founder and CTO · architecture, build and operations

fig.4 · one order, end to endsimulated

One order, five stations: customer pays, the kitchen's own system, ready, dispatched, delivered.

A simulated order moving through the platform. Orders land on the kitchen's own system, so the counter needs no extra tablet.
Design and make · Nov 2025 to May 2026

Team Bath Drones · aerodynamics lead

Leading the aerodynamics of a competition UAV.

I run the aerodynamics on Team Bath Drones' design-and-make UAV: more lift for every unit of drag, on less structure.

  • 30% lift-to-drag gain from an aerofoil trade study, Clark-Y against NACA 0015.
  • 400g structural weight saving, delivered at the Critical Design Review.
  • Co-led the tadpole fuselage and its CAD, then presented the CDR to the judges.
  • aerodynamics
  • aerofoil trade study
  • CAD
  • CDR
  • Clark-Y
  • NACA 0015

Aerodynamics division lead · University design team

fig.5 · aerofoil trade studyillustrative flow

Clark-Y against NACA 0015: +30% lift-to-drag in the trade study.

Profiles come from the real section coordinates. Streamlines are illustrative, not CFD. Selected: Clark-Y · +30% lift-to-drag in the trade study.
Paid engagement · invited back for a second year

Fifth9 · applied AI · red team and build

Attacking an enterprise AI platform, then helping ship it.

An agentic AI platform heading into regulated enterprise has to survive a determined adversary first. I play that adversary, then help ship.

  • Red-teams the platform's agents end to end over MCP, plus acceptance testing before client demos.
  • ~30 pull requests merged to production, from security fixes to new capabilities.
  • One memory, four models: a user's context carried across four major AI models on a single key.
Details
  • Shipped an internal code-search tool that maps the whole codebase.
  • Invited back after a 2025 internship: a generative-AI product shipped in a five-day sprint and pitched to the directors.
  • AI red-teaming
  • LLM evaluation
  • agentic AI
  • MCP tooling
  • UAT
  • full-stack

Red team, UAT, MCP tooling and demo builds · Fifth9

Client work under NDA

fig.6 · red-team consoleillustrative run

Red-team console and a four-model memory graph, played on invented, generic probes.

A scripted loop of generic probes, showing the shape of the work: probe, guard, patch, retest.
In progress · final report due November 2026

Year 2 design · snowmobile drive shaft

From 20kW at the motor to 209Nm at the track.

An electric snowmobile puts 20kW through one shaft at -40°C. I own the loads, free-body diagrams and drive ratio, built as a live engine.

  • 3.28:1 down to 209Nm at 915RPM, sized for 50km/h on a 290mm track sprocket.
  • A symbolic loads engine: change one input and every reaction, moment and stress updates, with 39 automated checks.
  • Final design locked: 431 stainless at a safety factor of 6.7, 4.35kg on Fenner Taper Lock hubs.
Details
  • Critical section at the drive sprocket: 80.4Nm peak bending, bearing reactions of 590.6N and 1340.1N.
  • Bearings sized for a 5200h life, about 285 million revolutions at shaft speed.
  • machine design
  • FBD / SFD / BMD
  • bearing life
  • three.js
  • JavaScript
  • unit tests

Pair design project, ME22007 · my lane: loads, FBDs, drive ratio and shaft torque

fig.7 · track drive shaftdrag to orbit

Motor 3000RPM · 63.7Nm to 3.28:1 to Shaft 915RPM · 209Nm.

02 / More builds

Twenty shipped, and counting.

Hardware, client sites, a tutoring business, bots and data pipelines. Most started as an itch of my own, then got hardened until they were reliable.

  • HardwareCoursework · 76.6%

    Silent looping render of a 5:1 epicyclic gearbox: the sun gear turns the planet gears inside a fixed ring gear and the planet carrier turns slowly as output.

    Epicyclic gearbox

    A 5:1 planetary gearbox, modelled, driven and rendered in Autodesk Inventor.

    Drive-constrained from a projected base sketch so the gears always mesh, fitted with standard bearings and circlips (BS 292, ANSI B27.7M), and backed by GD&T drawings marked at 88%.

    • Autodesk Inventor
    • GD&T
    • drive constraints
    • rendering
  • HardwareMechatronics coursework

    Ultrasonic measurement system

    An Arduino ranging rig calibrated to 1.0mm RMSE against a 4.3mm datasheet spec.

    Moving-average filtering and least-squares calibration cut noise by 60%; I mapped the ±18° beam and built a variance-based motion-rejection alarm.

    • Arduino
    • MATLAB
    • signal processing
    • calibration
  • HardwareYear 1 team of 6

    Ping-pong launcher

    Authored the slider-crank concept the whole team's launcher was built on.

    Chosen over five alternatives on a weighted decision matrix, then taken through full Inventor CAD inside a 380×180×300mm envelope and an £85 budget.

    • mechanism design
    • Inventor
    • DFM
    • decision matrix
  • Web & clientFounder · live

    Locus Tutors

    I turned my own tutoring into an online maths agency, and wrote the software that runs it.

    Stripe save-card billing and a password-gated admin CRM I built from scratch, monthly PDF progress reports generated from lesson recordings, PostHog attribution on every channel.

    • Stripe
    • Vercel
    • PostHog
    • Apps Script
  • Web & clientDe facto CTO · live

    Neetu's Crafts

    A full storefront for a family craft business, run end to end.

    Edge-cached Vercel read path, an Apps Script and Sheets backend, UPI deep-link payments instead of a gateway and automatic order emails.

    • Vanilla JS
    • Vercel edge
    • Apps Script
    • UPI
  • Web & clientExternal client · live

    Ruby's Rasoi

    A live restaurant site, shipped for a real client.

    Built to a non-technical owner's brief on a platform I did not choose, then handed over something they can run.

    • client work
    • web
  • AutomationPrivate

    LocaBot

    Six Chrome sessions in parallel, grabbing tickets the instant they drop.

    Driven over the DevTools Protocol and coordinated with asyncio, with staggered polling to stay under rate limits.

    • Python
    • asyncio
    • Playwright
    • CDP
  • AutomationPrivate

    BaianaBot

    A ticket sniper that watches an API and checks out before a human can refresh.

    Polls a live REST API, holds a persistent session and drives a Stripe checkout through sandboxed iframes, backed by a 150+ test suite.

    • Python
    • Playwright
    • REST
    • Stripe
  • AutomationLive · private

    Earnings calendar

    A self-maintaining calendar of earnings dates for a 44-stock watchlist.

    Runs unattended on scheduled triggers and rewrites each event with the realised price reaction afterwards.

    • Apps Script
    • Finnhub
    • cron
  • AutomationLive

    TES resource engine

    Turns exam specifications and past papers into typeset worked-solution packs.

    Generated as LaTeX, compiled with tectonic and published to the TES marketplace.

    • LaTeX
    • tectonic
    • automation
  • AutomationLive · local

    WhatsApp scheduler

    Queue a message, and it sends itself at a sensible hour.

    A local web UI and a self-restarting background worker with retry and backoff, standard library only.

    • Python
    • launchd
  • DataPrivate

    Family portfolio tracker

    Nine accounts and 37 holdings in one self-reconciling dashboard.

    Pulls live positions from the Trading 212 API, handles multi-currency pricing and failed feeds, and reconciles to the penny.

    • Apps Script
    • Trading 212 API
    • GOOGLEFINANCE
  • DataPersonal project

    Poker tracker

    A session tracker with real variance analytics.

    Standard deviation, 95% confidence intervals, breakeven win rate and expected value per session, computed from the raw data.

    • Vanilla JS
    • Apps Script
    • statistics
  • DataPersonal project

    Pool tracker

    A scoring app for a custom multiplayer pool game.

    Per-turn undo, persistent state and an all-time leaderboard on a serverless Sheets store.

    • Vanilla JS
    • localStorage
    • Google Sheets
  • DataPrivate

    Cricket tracker

    Play-Cricket scorecards scraped into a clean stats database.

    A logged-in Playwright scraper with match-level de-duplication, which also migrated an old manual dataset across.

    • Python
    • Playwright
    • Google Sheets
Aaditya Kalwani

About

Second-year engineer, compulsive builder.

I am Aadi, a second-year Aerospace Engineering student at Bath with a First from Year 1 (73.8%). The fastest way I learn a system is to ship one.

The thread through all of it: software and machine learning pointed at real engineering. A battery that cannot report its own charge, a shaft carrying 209Nm through slush, a kitchen that should not need a second tablet.

University
University of Bath · Aerospace Engineering with Placement
Placement window
2027/28
Based in
Bath, UK

Get in touch

Building something where this fits?

I am looking for a 2027/28 placement year in industry, and I am always up for a conversation about hard engineering, machine learning or anything on this page.

Download CV