Ahtesham Alvi

De Novo Protein Binder Design for IBD Therapeutics (IL-23)

University of Maryland — Independent Undergraduate Research (CMSC499A), advised by Prof. Fardina Alam

Undergraduate Researcher · March 2026 — Present

Python

Linux

Part of a two-student team designing de novo miniprotein binders against the IL-23R-binding face of the IL-23 p19 subunit (PDB 3DUH) — a target no prior de novo work had addressed, since existing miniprotein efforts aim at the IL-23 receptor rather than the cytokine itself.

• Built the project's computational scaffold and co-designed its first pipeline arm — a 9-stage RFdiffusion to ProteinMPNN to ColabFold workflow on SLURM HPC — then ran and analyzed the resulting screen, which produced no reliable binder across 972 designs (best ipSAE 0.30, none above the 0.60 reliable-interface threshold).
• Diagnosed the pessimistic single-sequence ColabFold signal from that screen as a scoring artifact rather than evidence of unfoldable designs: a roughly 56-point binder-pLDDT under-rating relative to AlphaFold initial-guess on identical backbones.
• The second pipeline arm, led by the primary author, applied interface-optimized hallucination (BindCraft) under the same scoring path — 42 of 50 designs above the 0.60 threshold, including five all-alpha-helical leads of 72-104 residues with Rosetta interface energies of -57 to -87 REU.
• First-authored a follow-on study superposing all 50 accepted designs into the experimental IL-23:IL-23R assembly: all 50 overlap the IL-23R-bound region, but 45 also clash with the obligate p40 subunit (0 to more than 3,500 atom pairs).
• Showed that constraint is tracked by none of nine interface-confidence metrics, corroborated by an independent sequence-only Boltz-2 analysis (Spearman rho = -0.60) — establishing assembly compatibility as a validation step distinct from interface confidence.

Temporal Robustness of Social Bot Detection

University of Maryland — Honors Research Seminar (CMSC396H)

Undergraduate Researcher · April 2026 — May 2026

Python

Scikit-learn

Pandas

Tested whether bot detection classifiers trained on historical Twitter/X data still work on modern bots, by training every model family on every era and measuring what happens when the eras do not match.

• Engineered account, behavioral, graph (PageRank, reciprocity, ego density), and node2vec features over four Twitter/X bot benchmarks — Cresci-2015, Cresci-2017, TwiBot-20, and TwiBot-22 — spanning 2015-2022 with up to 1M users and 88M tweets, normalized into a shared schema in DuckDB, NetworkX, and pandas.
• Trained six supervised classifiers (logistic regression, ridge, lasso, linear SVM, random forest, XGBoost) across all 16 train/test dataset pairs under a per-pair feature contract, yielding a 4x4 cross-era transfer matrix per classifier.
• Found every classifier attains near-perfect in-distribution AUC — up to 0.9955 — yet falls below 0.60 mean AUC across datasets, and that model family governs transfer behavior more than algorithm choice: the four linear models behave near-identically while the two tree-based models follow a distinct pattern.
• Attributed the generalization gap to dataset construction, labeling, and class balance alongside bot evolution, since models trained on newer data sometimes transferred worse to older data — indicating single-benchmark accuracy overstates real-world reliability.

Exoplanet Habitability Classification

University of Maryland, College Park

Independent Research · January 2026 — Present

Python

Scikit-learn

Consolidated four astronomical catalogs into a single feature-rich dataset and trained a semi-supervised model to assign habitability classes to confirmed exoplanets that had none.

• Merged the NASA Exoplanet Archive, the Habitable Exoplanets Catalog, ESA Gaia DR3, and SIMBAD into one dataset, maximizing both feature count and sample size.
• Trained a semi-supervised three-class model (non-habitable / mesoplanet / psychroplanet) on roughly 5,600 labeled planets, then predicted a habitability class for roughly 600 unlabeled confirmed exoplanets.
• Handled severe class imbalance with imbalanced-learn resampling and gradient-boosted ensembles.

RoboRaptor — Articulated-Wing Gliding Robot

FIRE Research Internship (FIRE199), University of Maryland — advised by Dr. Lena Johnson

Undergraduate Research Intern · May 2024 — July 2024

Arduino

C++

AutoCAD

Led a three-person team building a bird-inspired tethered gliding robot with articulating flapping wings, instrumented to characterize wind and atmospheric effects on flapping and gliding flight.

• Advanced the design across four prototypes benchmarked against house sparrow, peregrine falcon, and golden eagle wing geometry, raising the model's aspect ratio from 2.92 to 6.24.
• Replaced wooden dowel spars with carbon fiber and aluminum after in-flight structural failures.
• Wrote Arduino firmware for servo wing-torque control and SD-card flight telemetry, integrating an onboard wind sensor on a 3D-printed electronics plate.
• Achieved sustained flight in winds above 7 mph with in-flight wing actuation and directional control.

SeaTerp — Bioinspired Turtle Robot for Chesapeake Bay Monitoring

FIRE Bio-Inspired Robotics (FIRE298), University of Maryland — advised by Dr. Lena Johnson

Undergraduate Researcher · August 2024 — December 2024

Arduino

C++

Autodesk

Co-developed a tethered, sea-turtle-inspired underwater robot to survey dissolved-oxygen levels in the Chesapeake Bay dead zone, where nutrient runoff drives algal blooms that strip oxygen from the water column.

• Built the locomotion and sensing stack on an Arduino Uno driving six DC motors through two Adafruit FeatherWing controllers — crankshaft-actuated flippers, twin rear propellers, and a two-motor buoyancy bladder — instrumented with a dissolved-solids sensor and an onboard camera.
• Iterated the hull across successive CAD and 3D-printed revisions, moving from a rounded profile to a flatter ballast form for hydrodynamic stability.
• Validated subsystems over 20 bench trials spanning buoyancy control, dissolved-solids sensing at three water qualities, crankshaft actuation, and differential propulsion.
• The robot achieved 360-degree maneuvering and sampled water quality successfully, with an underwater trial ending early on hull ingress.