The portfolio
Browse the work by the problem it addresses. Each project note explains my contribution, the technical decisions, the evidence and the limits.
Simulation & optimization
I built a screenshot-to-plan workspace for a grid puzzle: review the board, search within a money limit, and inspect the result one turn at a time.
Working planner with source kept private. Screenshot interpretation needs review; scores are model predictions, not an observed Roblox result or proof of optimality.
Healthcare software
I built a patient-owned health record that connects clinical work across facilities, with consent enforced where the data lives.
Product in development. Core clinical workflows are implemented; scheduling is in progress. The portfolio interaction uses synthetic information.
Language-model security
I tested how poisoned source material changes a model’s answers, then measured what the defenses broke along the way.
Scoped research study at UW–Madison with public code and recorded results. The defense is not deployment-ready.
Agent systems
I built a writing workflow that traces claims to evidence, then hands the draft to independent reviewers with permission to reject it.
Working personal tool with Claude Code and Codex adapters. The source repository is private.
Knowledge systems
I built a research library that turns saved material into practical guides while keeping the path back to every source.
Live personal project. The public library is a curated static snapshot; private intake and write operations stay local.
Open-source contributions
I trace small failures in shared tools back to their boundary: an oversized schema, a lingering environment flag, or metadata that never reaches its consumer.
Submitted upstream pull requests, open at the November 4, 2026 review. These are proposed fixes, not claims of merged releases.
Language equity
I fine-tuned a multilingual transformer and used uncertainty sampling to study topic classification with limited labeled data.
Completed academic research project. Results depend on the corpus, augmentation and script conditions.
Published security research
I co-authored work on ransomware execution provenance, helping turn sandbox observations into reusable research evidence.
Published team research. The repository documents the collection pipeline; the dataset is available separately.
ML systems
I served an open-weight language model on a 6 GB laptop GPU and measured the tradeoffs between precision, throughput, latency and answer quality.
Reproducible personal benchmark on one laptop. The quantized weights come from the model publisher; I did not quantize them myself.
Distributed systems
I rebuilt a property lookup service with a cache tier, replicated storage, bounded failover and three implementations of the same contract.
Working personal systems project with public code and reproducible scripts. It is not a claimed production service.
Language-model security
I extended an existing red-teaming workflow and explored how the search and evaluator affected its results.
Course research using a fork of TAP/PAIR. Results were weak or inconclusive; this is not a measured safety-breakthrough claim.
Agent systems
I built a voice-workflow prototype that connects grounded research, verified pitches and a deterministic conversation controller.
Prototype. The documented version has not placed real phone calls. Discovery and voice components run locally; some writing and verification use cloud models.
Computer vision
I adapted a pretrained geospatial segmentation workflow to imagery of dense residential areas in Lahore.
LUMS course project. Uses pretrained SpaceNet weights; no from-scratch training or quantitative IoU/F1 evaluation is claimed.
Published research
I led an early study of social-media sentiment following the killing of Priyantha Kumara Diyawadana, presented as an ASONAM poster.
IEEE/ACM ASONAM 2022 poster. Sentiment analysis is not the same as a validated extremist-content classifier.
Machine-learning robustness
I trained a sentiment classifier and tested its sensitivity to small, constrained token substitutions.
LUMS course project from 2023. The result concerns the trained classifier and its experimental setup, not all sentiment models.
Computer vision
I worked on a mask/no-mask detector and projected detections from three camera views onto a common top-view plane.
LUMS computer-vision course project. Recorded detection scores used training images; they do not establish held-out accuracy. Tracking and duplicate removal were not implemented.
Engineering & leadership
I led the electronics and systems sub-team for a fixed-wing UAV competition entry.
Team competition project, March–July 2021. My documented scope was component selection, autopilot-software selection and circuit design.