Shubhan Mital
I am a Master’s student in Electrical & Computer Engineering at UC San Diego, specializing in Machine Learning and Data Science. I work at the intersection of AI, computer vision, and energy systems — drawn less to using models off the shelf and more to the math that makes them work.
At Yuanyuan Lab (UCSD), I research energy-aware scheduling for electric construction vehicles — more in the Research section below.
Experience
- •Data Scientist, Piramal Pharma — built ML forecasting pipelines, cutting processing time by 60%.
- •Data Science Intern, Piramal Pharma — automated reporting pipelines, lifting data accuracy to 90%.
- •SDE Intern, Amazon — built data-validation tooling that helped prevent up to $5M in losses.
Projects
- •SolarInvestAgent — agentic AI + MILP optimizer for solar/battery sizing.
- •StayTuned — multimodal vision + audio attention detection for classrooms.
- •Renewable Storage Optimization — deep learning models cutting VRFB power-loss error.
Outside of work, I enjoy analyzing Formula 1 strategies, football tactics, and cricket scorecards, going on hikes, and asking “why” a few too many times—both in coding and in life.
Education
UC San Diego
MS Electrical and Computer Engineering
Specialization in MLDS
BITS Pilani
MSc Mathematics
BE Electrical and Electronic Engineering
Technical Arsenal
GenAI & Deep Learning
- Agentic AI & LLMs
- Transformers (BERT/GPT)
- PyTorch & TensorFlow
- Hugging Face & Keras
- Computer Vision (YOLO)
- Multimodal AI (MediaPipe, Whisper)
Robotics & Embedded
- ROS 2 (Robot Operating System)
- OpenCV & Path Planning
- Sensor Integration (LiDAR)
- Embedded C
Cloud & Tools
- AWS (Lambda, S3)
- Azure (Databricks, ADF)
- Snowflake
- Docker & Git
- Excel & PowerApps
Languages
- Python
- C++ & C
- SQL & NoSQL
- MATLAB & Bash
- Java
Data Science Core
- Pandas
- Numpy
- Scikit-learn
- Predictive Modeling
- NLP
- Bayesian Inference
- Optimization (MILP/MPC)
Web & Full Stack
- React
- Node.js
- REST APIs
- Tailwind CSS
- Spring Framework
Yuanyuan Lab Research
Graduate Researcher • Smart Grids & Machine Learning
Oct '25 - Ongoing
- The problem going in: mobile construction equipment burns through battery fast, and nobody could say which part of the job was actually draining it. So I built a pipeline — YOLOv8 for detection, DeepSORT for tracking, a physics-constrained state machine, and a transfer-learned 3D ResNet-18 on top — that watches 34+ hours of raw jobsite footage (annotated via CVAT) and turns it into a minute-by-minute energy profile for every sub-activity an excavator performs.
- That pipeline now recognizes what the excavator is doing with 88.81% accuracy across 5 activity classes, up from an 86.7% benchmark that only handled 3. I layered in Bayesian regression to estimate battery state when the sensors go quiet, and built tooling to keep the training data balanced instead of hand-curating clips one by one.
- The current chapter is scheduling: comparing a deterministic MPC against a stochastic one over a rolling 24-hour window to figure out when and how to charge these machines without spiking demand charges or the grid's carbon footprint.
- Took all of this to the Jacobs School of Engineering Research Expo as a poster — turns out explaining certainty-equivalent MPC to someone holding a free taco is a great test of whether you actually understand it.
Industry Experience
Piramal Pharma Ltd. & Amazon Development Center
Data Scientist @ Piramal Pharma Ltd. Sept '24 - Sept '25
- Owned the full lifecycle of a forecasting model built to answer a genuinely hard supply-chain question 24 months out — which inventory would expire or run out before anyone noticed — from feature engineering all the way to production.
- Went down the pipeline rabbit hole too: rebuilt 4+ of them on Azure Databricks, ADF, and Snowflake, cutting processing time by 60% — which mostly meant far fewer 2am Slack pings about broken jobs.
- Turned all of that pipeline output into Qlik Sense dashboards that 20+ teams actually opened and used, instead of another report nobody clicks on.
- Automated pricing workflows and shipped NLP-driven chatbots that cut turnaround time by 70% — the best kind of automation, the kind where a human never has to repeat the same task twice.
Data Science Intern @ Piramal Pharma Ltd. Jan '24 - Jun '24
- Shipped 3 analytics and reporting tools end-to-end — from the first stakeholder conversation to something people actually opened every week — landing a 20% efficiency bump across the business units using them.
- Automated a pile of price scrapers and shipment trackers that used to be done by hand, pushing efficiency up 50% and data accuracy to 90%. Turns out most "data quality issues" are just tired humans copy-pasting numbers at 6pm.
SDE Intern @ Amazon Development Center July '23 - Dec '23
- Built data validation and pipeline auditing tools that sat quietly behind production UIs catching discrepancies across 5+ downstream pipelines — the kind of tool nobody notices until it saves the company up to $5 million.
- Got 2 critical pipelines onto real CI/CD with proper test coverage, shaving 15 hours off every release cycle — fewer manual checks, fewer "wait, did that actually work?" moments for the team.
Featured Projects
Deep Learning • Optimization • Multimodal AI
MediaPipe • TensorFlow • PyTorch • Whisper • ESP32
Engineered a real-time facial attention classifier using a 10-feature geometric BiLSTM (EAR, MAR, 3D PnP head pose), achieving a macro-F1 of 0.61 on the DAiSEE benchmark across 4 engagement-level classes. Fused facial predictions with a Whisper-based ~75-feature audio model so the system reads tone and pacing alongside posture and gaze, then deployed live multimodal inference to an ESP32 OLED display. Everything runs on-device — no video leaves the room — so a teacher gets a quiet, privacy-first read on whether the class is actually engaged, not just quiet.
LSTM • ANN • PyTorch • VRFB
Developed a deep learning framework with Dr. Ankur Bhattacharjee to predict and minimize power losses in Vanadium Redox Flow Batteries (VRFBs), a key technology for large-scale renewable energy storage. Modeled the non-linear relationships between electrolyte flow rates, state-of-charge, and power loss using ANN, LSTM, and RNN architectures in PyTorch, benchmarking all three against physics-based baselines to see where deep learning actually earns its keep. The best models achieved a 40–70% improvement in prediction accuracy, enabling tighter flow-rate control, less wasted energy, and meaningfully cheaper VRFB operation at grid scale.
LangGraph • MILP • HiGHS • Python
Built a hybrid agentic decision-support system for residential solar PV and battery storage sizing, combining a MILP optimizer (HiGHS) with LLM orchestration via LangGraph. Rather than trusting an LLM to reason about numbers, every hard decision — panel count, battery sizing, CAPEX, NPV, payback — gets delegated to a constrained solver grounded in real SDG&E tariffs, live irradiance data, and years of household consumption profiles. A three-stage validation loop (syntactic → schema → logical) catches anything the model gets wrong before it reaches the user. Across 20 benchmark scenarios, the agent cuts panel-count RMSE by 13.8× and CAPEX error by 2.3× versus an LLM-only baseline.
Hackathons
Building fast, under pressure, with strangers
Reboot the Earth is a global hackathon built around one premise: climate problems are too big and too urgent to leave to slow-moving institutions, so hand them to a room full of developers with a deadline instead. Our team built SAFE across the event, working through the entire pipeline — problem framing, data, and a working prototype — in a single sprint alongside teams from around the world tackling everything from disaster response to conservation.
It was less about walking away with a trophy and more about the format itself: no time to overthink architecture decisions, no room for perfectionism, just fast iteration with people you'd met a few hours earlier. That kind of constraint tends to sharpen your instincts faster than any amount of planning would, and it's part of why I keep coming back to hackathons even outside of climate-specific ones.
BrailleAI came out of a simple, uncomfortable observation: most assistive tech for the deafblind community is one-directional — it helps them receive information, but not send it back easily. We built a two-way interface instead: braille input gets converted to natural speech output, and incoming speech gets transcribed and routed back as braille or vibration patterns, so an actual conversation can happen instead of information flowing just one way.
We used Orkes to orchestrate the multi-step AI pipeline stitching speech, text, and braille together in real time, which ended up being the part of the build I'm proudest of — genuinely useful orchestration, not just tacked on for the sponsor track. The project won Best Use of Orkes among all submissions at the UC Berkeley AI Hackathon, and it's still the project I bring up first when people ask what I've built outside of coursework.