I’m drawn to hard problems in emerging markets.
I do academic research and lead AggieQuant.
Industrial Engineering
Brown Foundation Scholar with a full-ride merit scholarship. Minors in Statistics and Mathematics.
Brown Foundation scholarship awarded in 2025.
Previously, William Fremd High School in Palatine, Illinois. Now based in College Station, Texas.
Python, C++, Java, SQL, PyTorch, FastAPI
Full résuméEckman Lab · Texas A&M Operations Research
Researching uncertainty around noisy Pareto fronts through Gaussian-process simulations and mixed-integer optimization. Continued through an NSF REU in summer 2026.
95% coverage target, with 100+ posterior draws and 100+ Monte Carlo trials.
Pareto frontiers can look more certain than they really are, especially when the underlying objectives are noisy.
The workflow fits GP surrogates, draws posterior fronts, then solves a PuLP/CBC MILP for minimum-area confidence bands.
Sample fronts across noise regimes and kernel assumptions.
Use interval-selection variables, coverage constraints, and contiguous confidence regions.
Compare coverage and front sensitivity across repeated draws.
Python / OR / GP / MILP
Research plot / Eckman LabHackMIT
Built a live karaoke scorer using Spotify and Whisper, with a Flask backend and Next.js interface.
Blind Karaoke has one clean interaction: try to sing from memory, then see what happened. The result has to feel instant and easy to read.
The project connects audio capture, transcription, scoring, and feedback while keeping the machinery out of the way.
Record the attempt with minimal friction.
Compare remembered lyrics against the target.
Return a quick result that invites another run.
Flask / Next.js / Spotify API / Whisper
RepositoryFounder & President
Established two technical tracks and seven firm relationships, with research pods, workshops, and student onboarding.
Firm relationships include Jane Street, HRT, Citadel Securities, and PEAK6.
AggieQuant is the community project: events, resources, project work, and the small habits that help a curious group keep moving.
The hard part was not making more announcements. It was making it clear what was happening, where to start, and how people could contribute without hidden context.
Make onboarding and project entry points easy to find.
Keep talks, notes, and outputs visible outside meetings.
Use small recurring systems to keep momentum.
Quant / Ops / Community
AggieQuantPriced eight contracts across four live MarioKart races.
Built trading strategies for a simulated commodities exchange, competing among 30,703 players.
The Aggies finished #54 in America and #56 algorithmic, with 487,287 XIRECs.
Software development
Built workflows connecting company, supplier, and shipment data to source evidence. Merged 116 PRs in roughly 10 weeks.
Authored 143 pull requests across APIs, dashboards, and agent workflows.
Import research rarely begins with one clean table. Company identities, supplier relationships, shipment records, and market context live at different levels and must be reconciled before a conclusion is useful.
I built retrieval workflows, organized entity and shipment context, and exposed focused MCP tools so follow-up questions reuse the same evidence trail.
Normalize company, supplier, and shipment evidence for analysis.
Expose focused retrieval and analysis without hiding source evidence.
Carry context across comparisons and follow-up questions.
MCP / Data / Agents
Cambridge TRACE Lab
Prototyped Markdown-configured controls for model tone, refusals, monitoring, and usage norms.
Texas A&M InfoLab
Unified INSPIRED, ReDial, and ConvApparel behind a shared Python interface for conversational recommendation research.
Each conversational recommendation dataset ships with its own parser, fields, and vocabulary.
SpokenCRS provides one access layer for turns, item context, labels, and metadata so datasets are easier to query, validate, and compare.
Normalize heterogeneous dialogue datasets into common CRSDataFrame objects.
Represent user and system turns, item context, labels, and dataset-specific metadata.
Use loaders and notebook spot checks before model comparison and downstream experiments.
Python / NLP / Neo4j / PyTorch
Repository / InfoLabDekalb Capital Management · Risk Management Analyst
Contributed beta analytics, portfolio models, and trade-risk workflows to evaluate exposure, concentration, and position sizing.
Three contributions. Private repository.
The work sits close to the trading loop: understand the portfolio, measure exposure, and make each review easier to inspect.
The analytics support decisions around beta, sizing, concentration, and market sensitivity.
Estimate and review market sensitivity in a way that can be checked and reused.
Summarize exposure, concentration, and sizing context without hiding the assumptions.
Keep outputs simple enough to revisit when market conditions change.
Python / Beta / Portfolio Risk
Machine Learning Researcher
Investigating self-conditioning in diffusion language models and combining latent refinement with self-speculative decoding for more efficient inference.
Comparing hidden-state and vocabulary-space representations, including their effects on denoising convergence, output-projection cost, and memory overhead.
Assessing generation quality, refinement steps, and accepted tokens per full model pass.
Quantitative Analyst
Developing signals, backtests, and portfolio-risk workflows for a $6M AUM student-managed hedge fund.
Evaluating factor exposures across long/short equity and directional macro strategies. Translating investment hypotheses into reproducible backtests, scenario analyses, and risk cases before capital allocation.
Discretionary Trader Intern
Selected for a 12-week paid trading program with a structured $50K capital allocation, mentorship, and weekly strategy and risk reviews.
Researching catalyst-driven setups across futures, options, and leveraged equities. Maintaining IBKR trade logs with the mispriced variable, catalyst, counter-thesis, assumptions, instrument, position size, and thesis-based exit.
Python / PyTorch / Recurrent PPO
Built a simulated market to study how recurrent PPO learns adverse-selection risk from quotes and fills. Compared policies with Bayesian references and tested whether readable internal signals translate into better decisions.
Can a recurrent market maker infer the risk of trading with informed customers from its own quotes and observed executions?
Built a simulator separating informed trading from uninformed directional demand, with a Bayesian filter over 12 hidden market regimes. Compared recurrent, feedforward, recent-history, and belief-input PPO policies using held-out episodes and multiple training seeds.
In an audit of five frozen policies, quote-center choice accounted for 81.5% of the mean one-step score gap against the Bayesian quoting reference. Better risk decoding did not consistently mean better quoting.
A nonlinear recent-history decoder also outperformed linear recurrent-state probes. The experiments do not establish an advantage unique to older recurrent memory or causal use of the decoded risk signals.
Published protocols, per-seed results, and reproducibility checks accompany the code. The market is simulated; these findings are not live trading returns.
Python / SQLite
Built maker-only execution across Kalshi and Polymarket US, connecting fair-value pricing to inventory limits and venue-state reconciliation.
Venue adapters, live gates, and SQLite shadow sessions connect the strategy to execution. Inventory limits and venue-state reconciliation determine whether a quote becomes a position.
At HUTC, I priced eight contracts across four live MarioKart races and placed first among 48 teams.
Python / Kalshi API
Built a reproducible research pipeline for market ingestion, leak-safe features, walk-forward tests, calibration, and reports.
Prediction-market data gets messy quickly, and each manual step makes a backtest easier to misread.
The CLI ingests resolved markets, caches raw files, normalizes schemas, builds leak-safe features, runs walk-forward tests, and leaves an auditable report.
Keep raw JSON beside normalized tables for a clean audit trail.
Use time-aware features and splits before trusting model lift.
Compare priors, LR, microstructure, calibration, and ensembles in one loop.
Python / Kalshi API / Backtests
React / Flask / SQLite
Built browser-based SQL practice with target outputs, query attempts, and server-side checks. A tight feedback loop makes it easy to try again.
Built for TAMU Datathon Lite.
SQL Excavator makes database practice feel more tactile. You see a target, write a query, and get feedback fast enough to keep trying things.
The interface borrows the focus of a terminal while adding just enough structure to make the exercise approachable.
Show the expected output before the user writes SQL.
Compare submitted queries against expected results.
Keep attempts fast so learning stays active.
React / Flask / SQLite / sql.js
JavaScript / Product design
Designed a focused workspace for timed sessions, lightweight planning, and daily review. The next action stays clear, and the interface stays quiet.
Aevon treats focus as a loop: choose a task, start, finish, review, and carry forward what matters.
The interface stays quiet because this kind of tool has to survive repeated use, not just look sharp once.
Make starting and ending focus blocks quick.
Summarize progress so reflection stays light.
Keep controls calm, visible, and repeatable.
JS / UI / Product
NLP / Market data
Built a time-ordered pipeline for next-day equity classifiers, keeping future information out of the training process.
Built a time-ordered pipeline for next-day equity classifiers, keeping future information out of the training process.
Outside of work, I enjoy webnovels, manhwa, and manga. A few favorites:
