Swaroop Kumar Pal

Swaroop Kumar Pal

About

Imagine · Discover · Learn · Create · Iterate

What I do

I'm an engineer at Apple, where I research, build and ship estimation and on-device ML algorithms behind Location, Fitness and Wellness features running on millions of devices. Day to day, I architect algorithms around on-device sensors, build simulation platforms for rapid prototyping and evaluation, implement them to ship on hardware, and define test plans for data collection and verification. This is the exact kind of science and engineering role I've held throughout my career.

What drives me

The pursuit of engineering knowledge, both in breadth and in depth, and turning it into things that come alive has always been my North Star. And nothing beats the joy of seeing those products create a meaningful impact that is deeply personal to people.

That North Star has driven me to explore Estimation Science, AI/ML, Agentic AI, Robotics, Web Platforms, Embedded software, AR/VR, Sensor Fusion, and Computer Vision as a learner, builder, engineer, and researcher across industry-leading companies and institutions.

My desire to work across these domains is fueled by a curiosity about how complex systems are built from scratch, and an imagination for how these distinct domains can intersect to create something entirely novel.

Where I've been

In past lives I solo built an experiment tracking and evaluation platform for iPhone 11 GNSS positioning software at Intel, was founding engineer at Aireal (we built the world's first geospatial AR SDK), prototyped VR/AR infotainment for BMW's CES showcase, and shipped Jaguar's first in-car speech app. I hold an M.S. in Computer Science from UT Dallas. I'm currently at Stanford, "fine-tuning" my pre-trained computer science neurons through its AI Graduate Certificate.

What I do for fun

I dabble in technical art installations, algorithmic art, projection mapping, music synthesis, and robotics, always chasing the intersection of art and science.

Work

Apple 2019 – now

Location estimation & ML model optimization

  • Apple Watch Workout Distance — DRI for the fusion algorithm behind distance in millions of people's daily workouts.
  • Time in Daylight — DRI for the fusion algorithm powering the Health app feature.
  • Replay platform — DRI for the algorithm-development replay system: on-device code, built for scientists and engineers alike.
  • GNSS analytics — owner of the on-device and server analytics platform. Millions of privacy-preserved data points influence future hardware designs.
  • ML model optimization — led model tracking across 100+ teams; quantization, pruning and distillation saved 20% of system memory in iOS '24.
Intel 2017 – 19

iPhone 11 GNSS — experimented & validated on a platform I built and owned.

Aireal 2016 – 17

Founding engineer — the world's first geospatial AR SDK.

BMW 2016

VR/AR infotainment prototypes for the CES showcase.

FIVE Lab, UT Dallas 2015 – 16

Graduate researcher — VR apps for user studies; two published IEEE papers.

TCS · Jaguar Land Rover 2011 – 14

Jaguar's first in-car speech application.

Things I've Made

Engineermaxxing

A self-evolving education & research platform — agentic architecture, novel learning interfaces.
Expert-level learning and building for everyone.

2026 · In progress
Point and Pick

Diffusion policies for target-specific robot manipulation, on sim and real robots.

Honorable mention · Stanford
Claude Mambo

A voice-driven ambient computer — a projector wall that listens, answers and paints.

2026 · Voice × Wall
GAS — object detections overlaid on a Replica living-room frame
GAS

Visual navigation framework producing 2D/3D semantic maps from video.

2024 · Stanford
Silver Unicorn

An augmented companion robot, built at the MIT Media Lab hackathon.

1st prize · MIT
IoTxMR

A mixed-reality smart home on HoloLens — spatial UI across devices.

Featured in Digital Trends
Life is indeed a POMDP

A hybrid LLM + POMDP wellness agent inferring hidden user state from chat.

2025 · Stanford

From the Archive

2016
2015
2015
2015
2015
2015
2015
2015

Art × Science

Engineermaxxing

A self-evolving education and research platform to make expert-level learning and building accessible to everyone. Every morning it writes The Daily Delve — a narrated, personalized research brief — and grows a library of interactive lessons, paths and maps through agentic pipelines. Product design, agentic architecture and novel learning interfaces, all built nights and weekends.

Point and Pick

Point at an object and the robot picks it: diffusion policies for target-specific manipulation, conditioned on a pointed target. Built the MuJoCo simulation and Modal evaluation infrastructure spanning sim and a real UR arm with a fisheye gripper camera. Honorable mention among ~250 Stanford CS224R projects.

Claude Mambo

Hold a key, speak, and the bedroom wall answers. A local pipeline: on-device Whisper turns speech to text in ~0.1 s, a Haiku router reads the intent, and Claude specialists — ambient art, todos, info, cameras, system — run their tools and post scenes to a FastAPI WebSocket server, which a chromeless Chrome paints onto the projector. About a thousand lines end to end, offline except the model calls, with the same wall tools exposed to Claude Desktop over MCP.

What I'm saying in the demo

0:00"What time is it? Show me my to-do list. Cross off 'bring some coffee' on my to-do list."

0:28"I'm feeling very happy today. There is a party today — make it Hawaiian theme art."

0:57"How's the weather in San Francisco? Explain Kalman filter to a two-year-old."

1:26"What cameras do I have? Stream from the first camera."

1:47"Stream from the second camera. Take a picture. Show me my to-do list."

2:10"I have an Indian party today — make some art for Indian theme."

Sangeetmaxxing — LOOM

LOOM — listen, remix and own the world's living musical traditions. At its heart is Atlas, the infinite synthesizer: tap a vibe world — lo-fi, Afrobeat, Indian raga, gamelan — and a wheel of notes starts spinning; open the rack for euclidean drums and synth panels; record a take and publish it to the gallery. Behind it, Magenta RealTime 2 runs live on Apple Silicon — music generated at 25 Hz, streamed as 48 kHz PCM over a WebSocket into a gapless AudioWorklet — and a Perform mode conducts it with MediaPipe hand tracking: morph styles by moving your hand, raise it for drums and energy.

GAS — semantic maps from video

GAS team with the poster at the Stanford session

A visual navigation pipeline that turns monocular or RGB-D video into semantically labeled 2D and 3D maps — neural SLAM, a fine-tuned detector, SAM segmentation and geometric projection, evaluated on Replica (Stanford CS231n). GAS v2 rebuilds the pipeline with modern foundation models through agentic engineering: what can 30 minutes and 30 cents of compute do to a research paper?

Silver Unicorn

A tangible companion robot with mixed-reality, voice and touch interfaces, built in a weekend at MIT Media Lab's Reality, Virtually hackathon. 1st prize in the Getting Things Done and Doing Business category.

IoTxMR

A mixed-reality smart home: a HoloLens app giving a spatial, gaze-and-gesture UI to cross-platform IoT devices — an Android music player and an Arduino-driven fan and light — over TCP. Built at Microsoft's SF Holohacks hackathon and featured in Digital Trends; later re-imagined on Magic Leap.

Life is indeed a POMDP

A hybrid LLM + POMDP wellness agent that treats hidden well-being — body, mind, soul — as unobservable state, inferring beliefs from chat and choosing supportive actions with online forward search. Lives on as Eanors, a decision playground on the web.

Project Tango Hunt

A two-player augmented-reality treasure hunt built on Google's Project Tango with the Tango Unity SDK: one player hides virtual treasures around a real indoor space, the other follows clues and a trail to find them.

Pawn Shop in VR

A virtual pawn shop built in Unity 5 for Oculus Rift DK1, where users navigate and pick out glowing objects with Wii remotes — the environment behind the IEEE 3DUI head-tracking user study.

Surgical Robot in VR

A VR simulation of the da Vinci Xi surgical robot with interactive arm manipulation — Unity, Maya and motion capture, experienced on Oculus Rift DK1.

Google Cardboard

The pawn shop scene modeled in Maya, rebuilt as a Google Cardboard VR app in Unity — an expert-built environment in anyone's pocket.

Positional Tracking

An Android app integrating Unity and OpenCV: camera frames feed Lucas-Kanade optical flow so real-world steps become virtual motion, tested on Samsung Gear VR.

Kinect Projection Mapping

Dynamic projection mapping on real objects using a Kinect, with basic interactivity — an early experiment at the intersection of art and science.

IoT Smart Home

A model smart home where appliances and doors respond to face recognition — OpenCV vision with Arduino and AVR microcontrollers driving relays, motors and door locks over LAN.

Barrel Race

A motion-controlled Android game steered by accelerometer readings, drawn with SurfaceView and custom views.

Pick a soundscape to take a breather while you read.

The listening wall

A bedroom wall that listens — speak, and it paints ambient art, answers questions, or streams a camera. The full build lives in Claude Mambo under Things I've Made.

LeRobot picks, live

The Point-and-Pick rig at home: live detections, wrist-camera view, and a yellow arm sorting blocks off the desk.

Glyph rain

A generative rain of glyphs projected across the living-room wall.

Chair, remapped

A wingback chair dressed in projected stripes and patterns — furniture as canvas.

Guitar, projection-mapped

An acoustic guitar lit with spirals and stripes of light.

LeRobot learns to dance

An LLM choreographs a LeRobot arm — traveling waves through shoulder, elbow and wrist.

Dot-matrix mandalas

Dot-matrix mandalas and test patterns on the studio wall.

A wall of living frames

An installation wall where every frame animates under projection.

LEGO city, remapped

A LEGO diorama brought to weather and bloom by projection.

Cardboard companion

A googly-eyed cardboard robot that lives on the desk.

LLM pen plotter

Claude solves the math, then an AxiDraw draws the working — tool use over MCP, ink on paper.

Breadboard creature

Servos, speakers and an OLED face — a little machine with moods.