Hi, I'm Arjun Suryawanshi!
I study CIS for my Bachelors and Accelerated Master's at the University of Pennsylvania and am interested in Software Engineering and Quantitative roles, particularly involving Machine Learning. I enjoy solving hard problems and innovating new solutions, be it a custom computer vision model for classifying horse behavior or improved vehicle simulation tools, and strive to find such challenges to tackle.
Featured Projects!
FedCast
A central-bank policy-forecasting platform that infers Federal Reserve and ECB policymaker beliefs from unstructured communications data, then predicts upcoming rate decisions. Decomposes each meeting into five economic dimensions (employment, growth, inflation, financial conditions, uncertainty) with an importance weight and expected direction per dimension, combines them with realized macro data through estimated reaction functions, and forecasts the rate change (bps) and discrete action (hike/hold/cut).
- Team Project, repo private as private data was utilized. Created as part of Bridgewater Associates Immersion: AI for Innovation Hackathon Program.
- Built a data pipeline ingesting and cleaning ~5,000 central-bank speeches and ~5,200 policy documents across the Fed and ECB into a Parquet feature backbone (20,000+ dimension-feature rows, 10,000+ macro/market control series, 416 policy decisions) with automated QC reporting.
- Implemented a sequential Bayesian belief-updating model (pgmpy) that processes speeches chronologically — each posterior becomes the next prior — blending transcript, prior, and macro evidence into a five-dimension belief state.
- Estimated Ridge + multinomial-logistic reaction functions per institution, benchmarking macro-only, belief-only, and combined feature sets to quantify the marginal signal of transcript-derived beliefs; backtested over 2012–2025 on ~2,300 monthly snapshots, reaching up to 95% hike/hold/cut classification accuracy and ~5–12 bps rate-change MAE in-sample.
- Orchestrated multi-step reasoning with LangGraph/LangChain and live context retrieval via the Perplexity API, surfacing evidence traces, historical analogs, and skeptic flags for interpretability.
- Shipped a Dockerized FastAPI service with synchronous and async job-based forecast endpoints, progress polling, and on-disk forecast caching, plus an interactive JS dashboard featuring belief-history timelines and a rotatable 3D policy-dimension view.
Quasi-Static Vehicle Simulation
Quasi-Static Simulation for Penn Electric Racing: simulates a parameterized car and track to predicts lap times, energy use, and competition scores to systematically justify design decisions. Built in Python with Pydantic-validated subsystem models. My current contributions focus on the accumulator thermal simulation, the drivetrain efficiency model, and a CFD-backed aerodynamic map.
- Codebase private by team policy.
- Built a three-node transient cell thermal model (cell body, tab assembly, in-pack air) with temperature- and SOC-dependent internal resistance, entropic heat, Joule heat at the tab interconnects, Stefan-Boltzmann radiation, natural/forced convection (CFM-to-h correlation), and ideal-gas air density — integrated by forward-Euler over the QSS endurance current trace.
- Calibrated the thermal model against real Rev11 endurance telemetry via a coarse-to-fine 1-D sweep on the pack heat-rejection parameter, using energy-preserving RMS-vs-mean current resampling (RMS for heating, mean for SOC).
- Implemented a motor efficiency model as a 2-D tensor-product polynomial LUT (least-squares fit over RPM x torque CFD/dyno data, normalized inputs, Horner-method evaluation with out-of-bounds distance penalty), and inverted the implicit DC-bus power constraint tau*omega/eta(rpm,tau) <= P_limit by per-RPM bisection into a precomputed power-limited-torque table for O(1) runtime lookup, with automatic rebuild on parameter changes.
- Modeled the three-region motor torque curve (flat / field-weakening / cutoff), LUT-aware regenerative braking and coasting with RWD-vs-AWD motor-count conditioning, AWD 80 kW system power clipping, and added voltage derating to couple pack state into the available power limit.
- Replaced constant Cl/Cd aero coefficients with a CFD-backed aerodynamic map: a separable pitch x roll multiplier model for downforce, drag, and center-of-pressure, with body pitch/roll derived per simulation step from heave-spring and roll-stiffness suspension physics. Added attitude-driven center-of-pressure migration (front/rear downforce split) and a logarithmic radius-downforce factor for low-speed corners; validated with a Shapley-value decomposition of each effect's contribution to competition points.