UVA Darden School of Business

• Engineered an automated data extraction pipeline to parse unstructured BDC liabilities and joint venture disclosures from SEC 10-K and 10-Q filings, transforming complex regulatory disclosures into structured research datasets in collaboration with a Darden professor and a Federal Reserve economist.
• Designed and deployed a fuzzy name-matching algorithm across 7 TB of L2 voter registration data to match against hedge fund managers, resolving complex naming variations and edge cases with high precision for the Senior Associate Dean of ORS.
• Architected a data mining pipeline to extract, aggregate, and map multi-tier corporate hierarchies and parent company ownership records across diverse consumer brands.
• Developed a custom GPT-powered internal knowledge base enabling Darden faculty to query research profiles, publications, and ongoing projects, improving cross-faculty research awareness.

UVA Engineering Link Lab

• Analyzed large-scale EV fleet telematics to engineer battery capacity features, using OLS regression and ridge plot analysis to classify vehicle degradation tiers and validate partial-charge-based estimation as the top predictor.

Continuous Thought Machines - Sakana AI

Maze loss batch vectorization (merged PR #25)

• Vectorized the auto-curriculum masking in CTM's maze_loss() with PyTorch prefix argmax, clamping and broadcasting, replacing a per-sample Python loop (O(B×R)) with O(R) tensor ops and shrinking the mask from 3D to 2D.
• Fixed an IndexError that crashed training whenever batch size was above 1, unblocking batched and distributed (DDP) training, and hardened lookahead boundary clamping and the upto=0 edge case.
• Validated numerical equivalence against the original loop, boundary cases and runtime, achieving up to a 10× speedup at B=64, R=100 with fewer allocations; merged upstream into Sakana AI's official repo.

PyTorch, LoRA/PEFT, Qwen3.8-27B, A100 GPUs, SGP4, SQLite GitHub

• Trained five federated rank-8 LoRA clients of Qwen3.8-27B on real A100-SXM4-80GB GPUs, one data shard each, with orbital contacts, shared-station contention and outages simulated rather than flown; a round is admitted only when download, local compute and upload all fit the contact window.
• In one selected outage scenario and seed that contact-window scheduler accepted 14 local steps against 8 for fixed scheduling at 87.6 versus 147.1 MiB of modeled traffic; over 216 paired CPU cases it accepted more steps in 114, tied in 102 and accepted fewer in none.
• Kept interrupted transfers resumable with SHA256 chunk verification, idempotent receipts and an atomic SQLite round ledger; in a bounded two-round GPU test, two injected coordinator exits reproduced the uninterrupted final adapter and validation loss exactly.

LangGraph, LangChain, LangSmith, Harbor, FastAPI GitHub

• Built an agent that investigates alerts and fixes failures on a multi-service FastAPI platform using 21 typed LangChain tools through LangGraph: 15 read-only tools to diagnose, and 6 mutating tools that run only after a deterministic policy engine approves the action and issues a single-use HMAC approval token; resolved an injected outage in about 90s, with every step traced in LangSmith.
• Kept cost and latency predictable with hard per-run budgets (15 model turns, 35 tool calls), 60s LLM timeouts with retries, a 600s incident timeout and 50 KB tool-output truncation, plus a benchmark harness logging wall-clock time, model turns and tool calls per trial (about 16s scripted, about 90s live).
• Proved fixes are real with a Harbor evaluation harness where an independent verifier grades actual outcomes and penalizes cheating; the agent scored a perfect 1.0 reward in a LangSmith cloud sandbox trial, the harness surfaced 2 latent failure modes that local testing missed, and 83 tests pass in GitHub Actions CI on Linux and Windows.

Triton, LoRA/PEFT, Qwen3.8-27B, vLLM, A100 GPUs GitHub

• Built an execution-verified synthetic data pipeline so the training data is proven correct: Grok 4.7 teacher-LLM workers write Triton GPU kernels, an AST screen checks each one, and an Apptainer GPU sandbox on A100s runs it against a PyTorch reference, keeping only the 1,894 kernels that match.
• Used that data to LoRA/PEFT fine-tune Qwen3.8-27B on a single A100 via Slurm, raising held-out pass rate from 44.8% to 98.2% (1,178 of 1,200 tasks); ablations with a 0.5B LoRA model (12.3%) and a from-scratch 29.5M Transformer (0%) showed the gain comes from pairing verified data with a strong pretrained base.
• Served the base and LoRA models with vLLM to confirm the fine-tuned model is practical to deploy, at p50 latency of 10.3s and 101.9 tokens/s.

University of Virginia

CGPA: 3.925 / 4.0

Available to start full-time: June 2027

JNTU, Hyderabad