Research — Jineshwar Nariani
Strategic Deception in Financial Markets
Honors Thesis · Commonwealth Honors College · Advised by Prof. Patrick Flaherty & Prof. Hedyeh Beyhaghi
Decoy Trading and the Limits of Adversarial Detection. Can an investor hide a real trade among decoy orders well enough to fool an adversary watching the market? I built a multi-agent reinforcement-learning system to find out, pitting the investor against four detectors: deep RL, Bayesian inference, maximum likelihood and an urgency-based heuristic. I expected the deep-RL adversary to dominate. It didn’t: against a naive investor, the simple structural Bayesian detector won with 99.5% detection. Once the investor was allowed to adapt, it drove every detector towards a near-random equilibrium. Next steps are limit order books and latency, and reframing the adversary as inverse RL under strategic resistance.
Multi-agent RL, Adversarial Detection, Market Microstructure, Bayesian Inference, Maximum Likelihood, Game Theory, Portfolio Allocation, Statistical Modelling, Signal Extraction
PRISM
NSF Research · Multi-Agent LLM Systems
Agentic research workflows over long documents for four very different groups — a gene lab, a biotech lab, a photonics lab and the campus library’s resource office. Role-specialised agents for retrieval, analysis and synthesis run inside LangGraph graphs that handle branching and retries, and each lab’s data comes in through its own MCP server, so a new lab needs a new tool server rather than new orchestration. Every answer is retrieval-grounded, so researchers can check it against the source. Retrieval and modelling choices ship only after winning feature-flagged benchmarks, with models evaluated end to end (including H2O AutoML) before they reach researchers.
Agentic Systems, LangGraph, MCP, RAG, Evals, Feature-flagged Experiments, H2O AutoML
Financial RL for Portfolio Optimisation
Manning CICS · Paper Proposal Accepted at IEEE IDS 2025
An architecture that combines multi-agent reinforcement learning with large language models to personalise portfolio optimisation. I backtested and validated trading-strategy models under uncertainty with PyTorch and scikit-learn, judging them on statistical metrics so they generalise across volatile, data-sparse markets. I also implemented adversarial inverse reinforcement learning (AIRL) to model adaptive market makers from incomplete and noisy data.
MARL, AIRL, LLMs, Portfolio Optimization, Backtesting, PyTorch, Scikit-learn
Tariff-Policy Forecasting
UMass Math & Stats Summer REU · 2025 · Co-Authored Paper
What did 22 years of U.S. tariff policy (2000–2022) actually do to trade flows? I cleaned and restructured large tariff, trade-flow and macroeconomic datasets, then measured the relationships with Pearson and Spearman correlation, mutual information, Granger causality tests and random forests. On top of that analysis I engineered a hybrid forecaster: variational mode decomposition splits each series into signals, which an LSTM combines with macroeconomic indicators. It beat LSTM, BiLSTM and GRU baselines by 27%. I co-authored the resulting paper and presented the findings to academic stakeholders and project leads.
VMD-LSTM, Granger Causality, Time Series, Mutual Information, Random Forest, Econometrics, Forecasting
Cross-Domain Causal Transfer via Kan Extensions
CMPSCI 692CT: Category Theory for AGI · UMass Amherst · Spring 2026
An empirical test of Kan-Do-Calculus, the category-theoretic account of transferring causal knowledge between domains. I extracted causal triples from three real corpora — Federal Reserve documents, SEC enforcement releases and PubMed papers — then predicted causal graphs for 20 held-out queries with left and right Kan extensions, soft versions of each, and a retrieval baseline, scored on seven evaluation dimensions. Left Kan beat Right Kan on every dimension, confirming the predicted coend/end asymmetry; domain proximity predicted whether transfer happened at all; and the soft variants showed that the bottleneck is vocabulary grounding, not selectivity. Run on the UMass Unity cluster.
Category Theory, Causal Inference, Kan Extensions, LLMs, Sentence Transformers, NetworkX, Evaluation Design, SLURM / HPC
Program Synthesis vs. LLMs
Undergraduate Research Volunteer · UMass Amherst · Dec 2023 – Jan 2024
Can a classical program synthesiser still compete with a large language model? I built program synthesisers with Rosette, a solver-aided language, and evaluated them against LLMs on the same programming tasks.
Program Synthesis, Rosette, Formal Methods, LLM Evaluation