I study Data Science at UC Berkeley, and I spent this past summer as a Software Engineer Intern at Amazon in Seattle working on a batch ML inference platform. 4.00 GPA, Dean's List every term, though the transcript is the least interesting thing here.
Most of what I actually know came from building. There are 23 finished projects behind me: agents, computer vision, search engines, ETL pipelines, dashboards. I built a grocery budget optimizer because my mom wanted one. I wrote a data ethics module for high school students because that curriculum did not exist yet.
What gets me going: the moment a messy pile of data starts answering questions. Agents that recover instead of crashing. Code a stranger can still read six months later.
What I am not: someone who ships a notebook and calls it a product. I would rather spend the extra day on the schema, the failure path, and the tests.
Away from the keyboard I volunteer and mentor, 136+ hours logged so far.
Data Science · Data Engineering · Machine Learning · AI and Agentic Engineering
Always up for collaborating on AI, ML, Data Science, Agentic AI or Data Engineering projects. My inbox is open.
Redesigned how a batch ML inference platform resolves model configuration, then migrated it live behind a fallback architecture so no customer saw the switch.
- Merged 5 model configuration stores into 2 schemas and removed 20+ hardcoded service dependencies. Model onboarding went from weeks to hours.
- Built a fallback path across 3 distributed service layers so the live migration could not take the platform down.
| Changes merged | Lines added | Packages shipped | Incidents |
|---|---|---|---|
| 44 | 61,565 | 19 | 0 |
Search 29,000+ ASOS products by text, image, or both at once. Fashion-tuned CLIP embeddings feed two FAISS indexes, fused with keyword search, so "like this jacket but in linen" matches the picture and the words.
| Stack | Python · FashionCLIP · FAISS · BM25 · FastAPI · Next.js · Docker |
| Scale | 29,000+ products, two vector indexes, text / image / combined queries |
| Search | Vector and keyword results merged by Reciprocal Rank Fusion, then reranked on parsed intent |
| Links | Repository · Live Demo · Video |
Why it works: no single method handles real shopper queries. Vector search misses exact brand and size words; keyword search misses visual intent. Running both and merging is what made results usable rather than merely relevant.
A paid AI travel platform where six LangGraph agents research, draft, critique and finalise an itinerary. Requests route across 10 LLM keys with health scoring, so one provider going down never reaches a paying user.
| Stack | Python · FastAPI · LangGraph · React · PostgreSQL · Stripe |
| Scale | 6-agent state machine, routing across 10 keys, 99.9% uptime |
| Product | Stripe Free / Pro tiers, usage tracking, webhooks, quota enforcement |
| Links | Repository · Live Demo · Video |
Why it works: agent demos are easy, agent products are not. The engineering went into what users never see: quota enforcement that survives a replayed webhook, routing that degrades instead of failing, and a critic agent that catches a bad generation before a customer does.
An ETL and analytics pipeline over 15.6M+ transaction records, built to answer a question the business could not previously ask: which vendors quietly tie up working capital?
| Stack | Python · SQL · Pandas · Power BI |
| Speed | Query time cut from 9 minutes to 44 seconds, roughly 12x |
| Finding | $3.7M of unsold inventory capital held by underperforming vendors |
| Links | Repository |
Why it works: the 12x mattered more than it sounds. At nine minutes a query, analysts asked one question a day. At forty-four seconds, they explored.
Dialogue summarisation on FLAN-T5 fine-tuned over 16,000+ SAMSum conversations, using a PyTorch training loop written from scratch with mixed precision and resumable checkpoints.
| Stack | Python · PyTorch · FLAN-T5 · FastAPI · Groq · React |
| Quality | ROUGE-1 = 43.53, sub-second inference, 35 tests passing |
| Links | Repository · Live Demo · Video |
Why it works: writing the loop by hand instead of using a prebuilt Trainer was the point. Checkpoint resumption and profile switching break when you cannot see the loop, and they are what let the model train on free compute that can be interrupted.
Predicts whether a Falcon 9 first stage lands successfully, then turns that into launch-cost economics. Collection, cleaning, exploration, mapping and modelling end to end.
| Stack | Python · Pandas · scikit-learn · Folium · Plotly Dash |
| Result | 94.4% accuracy, SVM selected from 4 classifiers under GridSearchCV |
| Impact | Quantified a $103M cost difference per launch based on stage recovery |
| Links | Repository |
Why it works: the modelling was the short part. The value came from reconciling an inconsistent public API against scraped launch tables, which is where the real errors lived.
A 2D fluid-dynamics solver written from first principles for UC Berkeley Physics 77, used to sweep airfoil shapes for the best lift-to-drag ratio.
| Stack | Python · NumPy · finite-difference methods |
| Result | Best lift-to-drag = 1.479, NACA 5315 at 0.1 m/s |
| Recognition | Charlene Conrad Liebau Library Prize, Honorable Mention. The only STEM paper among lower-division finalists from 51 applicants |
| Links | Repository |
Why it works: implementing the pressure coupling by hand rather than calling a solver library is why this one is here. It is where numerical stability stopped being a debugging problem and became a design constraint.
| Area | Level | Evidence |
|---|---|---|
| Agentic AI and multi-agent systems | Production | 6-agent LangGraph workflow serving paying users, failover across 10 providers |
| Search and retrieval (RAG) | Production | Vector plus keyword retrieval over 29K products, two FAISS indexes fused by rank fusion |
| Fine-tuning transformers | Proficient | FLAN-T5 on 16K dialogues, hand-written fp16 loop, ROUGE-1 43.53 |
| Classical ML and model selection | Proficient | GridSearchCV across SVM, Random Forest and CNN; 94.4% and 99% on two problems |
| Data engineering and ETL | Proficient | 15.6M-record pipeline, 12x query-time cut, production schema work at Amazon |
| Deployment and MLOps | Working | Dockerised services on HuggingFace, Vercel, Render and Railway; full test suites |
| Recognition | What it is |
|---|---|
| 4.00 GPA, Dean's List ×2 | UC Berkeley, B.A. Data Science, every graded term |
| Amazon Future Engineer Scholar | National scholarship that includes the Amazon internship |
| Greenhouse Scholar | Whole College Program, selected at a 1-in-1,780 rate |
| RSM US Foundation First Generation Scholar | 2026 cohort, 1 of 5 nationwide |
| Charlene Conrad Liebau Library Prize | Honorable Mention, only STEM paper among lower-division finalists from 51 |
| Valedictorian | Ranked #1 of 455, AP Scholar with Distinction |
Seven more honours and fellowships
| Recognition | What it is |
|---|---|
| MLT Ascend Scholar | Management Leadership for Tomorrow career fellowship |
| AI4ALL Ignite Fellow | Applied AI accelerator for underrepresented technologists |
| CAA Leadership Scholar | Cal Alumni Association multi-year leadership award |
| QuestBridge National College Match Finalist | Also a College Prep Scholar |
| Berkeley competitive prizes | Leslie Lipson · Elizabeth Mills Crothers · Dorothy Rosenberg · Lili Fabilli and Eric Hoffer |
| IMO Gold Medal | International Mathematics Olympiad, Level 1 |
| GFWC National 1st Place | National youth writing competition |
| Certification | Issuer | Issued | Credential |
|---|---|---|---|
| AWS Certified AI Practitioner | Amazon Web Services | Aug 2026 | view |
| Foundations of AI Engineering, Honors | CodePath | May 2026 | view |
| Berkeley Student Leadership Academy | UC Berkeley | Apr 2026 | view |
| Berkeley Changemaker | UC Berkeley | Dec 2025 | view |
| Deloitte Australia Data Analytics | Forage | Jul 2025 | view |
| Commonwealth Bank Data Science | Forage | Jul 2025 | view |
| Introduction to Data Analytics | IBM | Feb 2025 | verify |
| SQL for Data Science | UC Davis | Feb 2024 | verify |
| Python for Everybody, Specialization | University of Michigan | Dec 2023 | verify |
building:
- "Automation and personal assistant agents"
- "Computer vision projects"
- "Data science work, end to end"learning:
- "Data engineering"
- "Discrete mathematics"
- "NLP and large language models"
- "Agents and agentic coding"open_to:
- "Summer 2027 internships: Data Science, Data Engineering, ML, AI Engineering"
- "Collaborating on AI, ML, Data Science, Agentic AI and Data Engineering projects"Hiring, building something adjacent, or comparing notes on agent evaluation? I answer every message.
Richard Hamming