Quant & AI
case studies.

Reads an option chain and reports which quotes cannot all be true at once — below intrinsic, out of order in strike, spreads dearer than they can ever pay, broken butterfly convexity, broken put–call parity. Every check is model-free, so a violation is a statement about arithmetic rather than a model's opinion, and each finding names the cheapest explanation to check first: on a real chain it is a stale quote or a missing dividend, not free money. The implied volatility solver returns nothing rather than a number where a quote does not identify one. Zero dependencies, no network, exits non-zero. MIT.

Reconciles a purchase order against goods received against the supplier invoice and names one reason each line disagrees, covering the three ways a figure changes meaning in transit: identity, unit of measure, and period cut-off. Zero runtime dependencies and Decimal throughout, because the people who need it run it on machines where installing a library is a conversation with IT. Exits non-zero on exceptions so a pipeline can gate on a clean reconciliation. MIT.

A small exporter's customs-cleared shipping bill becomes a receivable they can discount on day one. When the foreign payment lands weeks later, a matcher ties the messy SWIFT message to the shipment — the LLM only proposes, rules re-verify, thresholds and humans decide — then the ledger repays the financier and pays the exporter the balance. Money never moves on-chain; the contracts only record who is owed what. Hackathon build.

Turns a phone into a vibration, acoustic and magnetic sensor for rotating machinery, scores each reading against a per-machine reference, and trains one fault classifier across three or more phones over Wi-Fi Direct with no cloud. A team build for the iQOO hackathon; I wrote the federated learning, anomaly scoring and peer-to-peer sync, about 60% of the code.

A reference engine for an Android layer that screens what is on screen, offline: watermark detection for AI-generated images and video frames with a local classifier fallback, and phishing scores for links in web pages, SMS and email. A cheap always-on triage pass decides what earns the heavier model; nothing leaves the device and the event log is encrypted. The Android and NPU binding is the next phase.

Extracts specs from messy industrial product text, then asserts every number into the Z3 theorem prover as physics and domain constraints. When the values cannot coexist — a bearing bore wider than its outer diameter, power that is not volts times amps — it returns the smallest contradicting set as a plain-language reason, with outlier, duplicate and drift checks on top. The LLM extracts; the solver decides. UNIHACK build.

How much real data does a self-consuming loop of tabular generative models need before it stops collapsing? Gaussian copula, CTGAN and TVAE on three public datasets, measured against closed-form theory. One finding: rank-based copulas never collapse in variance — the damage moves into dependence, support and downstream utility instead. Paper pending conference presentation.

Agent-based simulation (500 agents, 160 Monte Carlo runs) of systemic risk from homogeneous LLM traders: +30.3% drawdown amplification at full adoption. Proposed Computational Asymmetry Theory as a risk framework.

Nightly ML signal pipeline for NIFTY/BANKNIFTY: 6-model ensemble (HMM, GJR-GARCH, LightGBM, LSTM, meta-model) over 80+ features, with VaR limits and a strict paper-trading gate.

Forensic RAG engine answering investor questions over Indian listed-company filings, with citations to the exact filing, year, and page across a 250-filing corpus target.

Multi-agent pipeline orchestrating 6 AI agents from idea to finished paper; 88% retrieval accuracy over a 50K+ document corpus, 70% cut in research cycle time.

GNN-PPO routing for distributed quantum sensing networks: near-Heisenberg scaling (α=1.92±0.04) and +51% quantum Fisher information over the strongest baseline.

Hybrid quantum-classical cyclone intensity classifier using IQP-kernel QSVMs on IBTrACS storm data: 96.1% accuracy with noise-resilience and OOD analysis.

pnpm/TypeScript monorepo: Expo/React Native app, Fastify 5 + PostgreSQL API, shared Zod contracts typed end-to-end, and Cashfree e-mandate checkout.

Next.js financial index prediction app with real-time market data and Gemini-based forecasts; ML model over 500K+ data points with Monte Carlo simulation.