Four teams. One pattern: identify the friction, engineer the system, ship it at scale.
Built 8 Python agentic systems (Selenium, pandas, openpyxl, Outlook API) executing I-9 and immigration workflows across Workday, Equifax, and Sprintax.
Cut processing time from 1–3 hours to under 5 minutes per task, saving ~28 hrs/week and up to ~40 hrs/week during peak periods.
Built an Outlook-integrated email agent (Python + Microsoft Graph API + rule-based classification) that reads I-9/immigration inbox emails, classifies queries, pulls employee data, and replies without manual intervention.
Delivered ~$40K in annual savings and wrote documentation that lets non-technical staff run and troubleshoot the systems themselves.
OCT 2025 – PRESENTBuilt ingestion and validation pipelines in Python and R across field and sensor data sources, cutting manual cleaning time by ~30 hours per week.
Developed statistical agents that automatically flag missing values and anomalous sensor readings, catching 325+ data quality issues before they reached downstream models.
Set up computer vision preprocessing with torchvision to standardize and augment ecological image data, taking preprocessing from ~3 hours per batch to fully automated across 25,000+ images.
JAN 2026 – MAY 2026Designed and ran evaluation suites covering 200+ test scenarios to measure AI agent accuracy, response quality, and edge-case behavior.
Built automated LLM testing workflows with batched, parallelized execution, cutting manual evaluation time by ~60%.
Constructed benchmark datasets that surfaced failure modes feeding directly into model improvement cycles, improving total test coverage by 35%.
Presented evaluation findings and coverage gaps to the agent team in weekly syncs.
JUL 2025 – AUG 2025Built multi-granularity datasets (daily, weekly, monthly) from 5 years of raw time series data, engineering features to surface trend, seasonality, and autocorrelation.
Trained and benchmarked ARIMA, Prophet, XGBoost, and LSTM; XGBoost outperformed by 18% on RMSE (weekly granularity), later adopted into the team's production pipeline.
Shipped a resampling frequency analysis that cut the team's trial-and-error preprocessing time by ~5 hours per experiment.
MAY 2025 – AUG 2025Things I built on my own time because I wanted them to exist.
- Built a full-stack job-hunt SaaS: Sam aggregates roles from GitHub sources, ranks them by rule-based scoring (role fit, location, visa, company stage, recency), and tracks applications with recruiter outreach.
- Engineered apply-core as a universal form-fill engine with computer-vision fallback for unsupported portals; the same DOM-inspection and field-resolution code runs in both the Chrome extension (production) and Playwright (testing/CI).
- Architected Rosie's automation as server-orchestrated browser control: the server handles LLM planning, RemotePlanner directs the extension to fill forms in a real browser (cookies, SSO, CAPTCHA pauses, file uploads), session sync keeps state consistent across layers.
- Shipped to live beta with 50 active testers and 500+ waitlist signups, actively iterating on fill resilience, Workday expansion, and cross-device persistence.
- Built and operate a services business designing and shipping production agentic systems and custom web applications for local businesses and startups; own the full lifecycle from requirements mapping to deployment, monitoring, and ongoing iteration.
- Deployed 4+ production workflows handling invoicing, follow-ups, lead routing, inventory sync, and reporting, designed for reliable operation under real volume and edge cases, with error handling, logging, and observability built in from day one.
- Architect agentic workflows that automate operational tasks (triage, routing, communications) with guardrails and human oversight, integrating across existing client tools to eliminate manual re-keying and create single-source-of-truth dashboards.
- Operate with accountability for outcomes, not hype; measure success by reduced manual steps and operational friction, positioned for transition to recurring agentic automation retainers.
Things I built because I wanted to know if they'd actually work.
- Engineered 6 CUDA attention kernel variants, implementing shared-memory tiling, online-softmax fusion, vectorized memory access, asynchronous double buffering, and Tensor Core acceleration.
- Achieved a 57.9× simulated speedup over naive attention using Tensor Core WMMA kernels, identifying memory bottlenecks, register pressure, and GPU occupancy constraints.
- Built a counter-guided autotuning system that analyzes GPU performance metrics, selects kernel transformations, and validates optimization decisions against predicted hardware behavior.
- Validated 23 kernel configurations on a real NVIDIA A100, achieving 0.800 Spearman ranking correlation between simulated and measured performance, with the Tensor Core variant ranking fastest.
- Debugged and patched 5 GPU simulator defects involving Tensor Core emulation, warp shuffles, and instruction latency, enabling reliable correctness testing and architectural analysis.
- Built an end-to-end vision-language model compression pipeline, fine-tuning Qwen2.5-VL-3B on VQAv2 and evaluating quantization, LoRA, QLoRA, and knowledge distillation.
- Compressed a 7.51 GB model to 1.54 GB (4.9× smaller) while retaining 97.6% of baseline VQA accuracy through teacher-student distillation and GGUF quantization.
- Achieved 41.6 decode tokens/sec and 2.69-second end-to-end inference latency per image question on a 16 GB MacBook Air M3 using llama.cpp and Metal.
- Benchmarked GPTQ, AWQ, and GGUF across multiple precision levels, measuring accuracy, memory footprint, and inference speed to identify practical compression tradeoffs.
- Developed reproducible training and evaluation pipelines with cross-backend validation, automated reporting, and a fully local browser demo for image-based question answering.
- Built a distributed LLM training system using PyTorch DDP and FSDP2, benchmarking GPT models across 1, 2, and 4 NVIDIA H100 GPUs.
- Achieved 92.5% weak-scaling efficiency and 933K tokens/sec with DDP on 4 GPUs, measuring training throughput, synchronization overhead, and GPU utilization.
- Reduced per-GPU model-state memory by 75% using FSDP2 sharding, enabling training of 12.85B-parameter models, 4.85× larger than the largest tested DDP configuration.
- Optimized distributed training with bf16 mixed precision, fused AdamW, and NCCL collectives, improving FSDP training step performance by 1.8×.
- Developed reproducible GPU benchmarks, memory profiling, communication analysis, and correctness tests to identify bottlenecks in distributed training.
- Architected a full-stack evaluation harness for LangChain tool-using agents: trajectory interception (DAG capture of tool calls, args, outputs, latency), parallel multi-metric scoring, and regression detection against rolling history.
- Built a non-invasive LangChain tool wrapper that intercepts every agent decision point, recording it as a DAG node with reasoning traces, token counts, and parent-child links to reconstruct the full decision trajectory.
- Designed three concurrent scoring metrics to catch failures end-to-end answer checks miss: tool-call accuracy (Levenshtein distance), semantic similarity (sentence-transformer embeddings), and faithfulness (claim grounding).
- Implemented z-score regression detection against a rolling 30-run history, flagging quality degradation as small as 1.5σ before it cascades into production.
- Shipped SQLite persistence (four-table schema) plus a Streamlit dashboard with KPI strips, pass-rate bars, metric distributions, and full trajectory replay, all local, no external services.
- Built a six-object ontology with four link types from six synthetic datasets, then architected six agentic workflows grounded directly in it.
- Chose SQL-over-ontology retrieval instead of embeddings, since exact case_ID matches mattered more than semantic recall.
- Root-caused a grounding failure to an unscoped case_ID, fixed it, and hit 93% grounding accuracy on a 15-case held-out test set.
- Cut response latency 4.2s → 2.1s by trimming the input payload to deal-relevant fields only.
- Built a RAG chatbot with OpenAI + LangChain, handling ingestion, chunking, and FAISS vector embeddings from scratch.
- Cut irrelevant retrieval by 20% and latency from ~4s to under 1.5s through optimized chunking and prompt engineering.
- Deployed to 100+ users over a 2-week live beta with 98%+ uptime.
- Built computer-vision tag fallback reducing manual photo-matching time by 85%, enabling recovery of ~60% of items with missing/damaged NFC tags.
- Built a React + Firebase lost-and-found platform mapping NFC/QR tags to dynamic, Firestore-backed recovery pages.
- Built an admin dashboard with search, filters, pagination, record stats, and CSV export.
The stack, end to end.
Foundations.
Coursework: Data Structures & Algorithms, Computer Systems, Probability, Applied Statistics, Database Systems