Data Science | Statistics | @ Rutgers University, New Brunswick -- I build systems where models meet infrastructure. Python for modeling, Rust for the hot path, LLMs for reasoning.

Kernel-enforced cleanup on wrapper exit. Windows uses a Win32 Job Object with JOB_OBJECT_LIMIT_KILL_ON_JOB_CLOSE, Linux uses systemd-run --user --scope with cgroup.kill on kernel 5.14+ and falls back to setpgid + a bash trap on older kernels and WSL1. macOS is supported at a documented honesty ceiling using setpgid + a disowned out-of-process watchdog, pinned by a negative test that proves the failure mode rather than hides it. 78 tests across PowerShell and bash, CI matrix on all three platforms.

Reads a PDF section by section and streams structured notes over SSE from a local Ollama model, so there's no API key and nothing leaves the machine. Every note carries an anchor chip: click the card and the PDF scrolls to the matched passage and flashes it. The resolver tries an exact text match, falls back to fuzzy matching when the wording drifts, and degrades honestly to page-level anchoring with the chip telling you which happened. Built for paper: notes export to wide-margin or Cornell print layouts for copying out longhand.

15-60 second build demos in a TikTok-style vertical feed for shipped GitHub projects. 33 numbered Postgres migrations behind a Codex adversarial-review CI gate that blocks the PR if a reviewer's findings haven't been answered in the diff. GDPR Article-17 erasure is wired into automated PII drift checks that fail the build if personal data leaks into a new column.

Local prop EV engine covering projection modeling, multi-book line sweep, probability calibration with isotonic regression, 8-layer signal gating, historical backtesting against 9M+ odds snapshots, and CLV-tracked paper trading. FastRust is a companion Rust compute engine via PyO3 that handles the hot-path EV math with GIL-releasing batch parallelism, drop-in replaceable with the Python compute_ev(). The run's charts and verified numbers are public in courtside-showcase.

Wrote the rendering engine from scratch (no Phaser or Pixi) with A* pathfinding, 4-directional sprite animation, and map transitions. 10 NPCs each route to their own LLM provider (Anthropic, xAI, Ollama, or local subprocess). Agents consult each other mid-conversation through tool-call delegation, so the world's state is shaped by inter-agent negotiation, not scripted dialogue trees.

RedLine -- SEC filing change detection with local LLM analysis

Watches any SEC-reporting company on EDGAR. When a new 10-K or 10-Q drops, diffs each section against the prior filing at the sentence level using cosine similarity and greedy bipartite matching, flags material changes against 9 red-flag patterns, and runs risk scoring through a local Ollama model. No paid APIs, no cloud dependencies.

Python Ollama EDGAR cosine similarity bipartite matching

Also public: Crash (desktop marketplace where AI agents shop for tools and pay over Coinbase x402), spotify-cleaner (dry-run-first CLI that prunes your least-listened liked songs), and Cluely (capture-invisible meeting overlay for Windows).

Intuit TechWeek SMB Underwriting -- Explainable-ML underwriting challenge (NY Tech Week 2026, 2nd place)

Our team took 2nd place at Intuit's hackathon at Intuit HQ. We converted calibrated PD models (validation AUROC 0.740) into expected-NPV approval decisions for 13,306 applicants, forecast how the funded book defaults over 13 weeks, and answered 900 causal what-if queries with per-feature guardrails. The dataset was booby-trapped with selection bias and label leakage on purpose, and finding those traps was most of the scoring. I owned the shared feature pipeline and the approval-policy deliverable.

Python scikit-learn probability calibration causal inference survival analysis

StormLink -- Multi-agent disaster response dashboard (HackUSF 2026, Best Use of ElevenLabs)

Co-built a multi-agent platform that routes live NOAA and FEMA feeds through Google ADK agents over the A2A protocol, with ElevenLabs voice alerts surfacing material weather events in real time. Wrote the AgentPanel component that visualizes the routing pipeline with animated data packets and elapsed timers.

Google ADK A2A protocol Gemini 2.5 Flash ElevenLabs Next.js 16 FastAPI Leaflet

InternPilot -- Automated job application pipeline

Pipeline that scrapes 13 job sources (JSearch, Greenhouse, Lever, Indeed, LinkedIn, Handshake, and more), scores postings against a resume profile, auto-fills ATS forms via Playwright across Greenhouse, Workday, Lever, Ashby, iCIMS, and SmartRecruiters, and drafts cover letters through a local LLM. Currently tracking 127 target firms.

Python Playwright Ollama FastAPI APScheduler

nemotron-reasoning-challenge -- Autonomous fine-tuning pipeline

Daemon that loops through 10 stages: Claude-powered math CoT generation, QLoRA SFT via Axolotl, GRPO reinforcement learning, vLLM eval, failure diagnosis, and Kaggle submission. Deadline-aware strategy that shifted from exploration to exploitation as the June 2026 deadline approached.

PyTorch Axolotl trl GRPOTrainer vLLM Claude API