Projects

Helix

Built a self-correcting heart disease classifier — an autonomous ML agent that runs the entire training loop itself with no human in it. Given a prediction task on the UCI Heart Disease dataset (Cleveland, Hungary, Switzerland, Long Beach VA, ~920 records combined), the agent collects data, trains, evaluates itself, and corrects course until it beats the published benchmark (83.3% baseline, up to 86.9% best-published), with no human choosing the model, tuning hyperparameters, or deciding how to fix a bad result. It starts deliberately weak with only the smallest hospital source and has to diagnose on its own whether it's underperforming because it's data-starved, the data shape is wrong, or it picked the wrong model class — then pulls in another hospital source live, transforms data through Nexla, discovers and pays for an enrichment service through Zero.xyz mid-loop, or switches models, and retries. Every autonomous action — pulling a new data source, spending on enrichment, swapping models mid-loop — is gated through Pomerium with a spend ceiling, action allowlist, and full audit log, making it possible to demo live and prove after the fact that the autonomy stayed inside real guardrails.

AI Agents Pomerium Zero.xyz

SubStrata

Engineered a recursive self-improvement pipeline for NASA-oriented land cover classification, boosting model accuracy from 42% to 91% through iterative error correction. Built a dual-model feedback loop where Gemma audits Gemini 3.5 Flash's classification output against ground truth labels, pushing flagged discrepancies to a persistent memory layer that compounds accuracy gains over time. Integrated Google Earth Engine's API to source and process Sentinel-2 and Dynamic World satellite imagery for training and validation.

Python Gemini Google Earth Engine

Cascade

Formulated N-2 cascading failure scenarios in power grids as a QUBO problem, achieving 100% contingency coverage on real IBM quantum hardware. Engineered a quantum optimization pipeline that identified failure-containment strategies ~8x better than a greedy baseline, landing within 22.5% of the true optimum — deployed and validated on live IBM quantum hardware.

Quantum Computing QUBO IBM Quantum

PathFinder

Built a web-only navigation aid for blind and low-vision users, running entirely in-browser on WebGPU with no specialized hardware required. Fused on-device object detection with depth estimation to deliver real-time obstacle warnings via directional haptics, audio cues, and spoken alerts with zero cloud latency. Implemented a live-scene query feature integrating Bright Data web search to identify unfamiliar objects, and a RunPod pipeline for continuous retraining without disrupting on-device performance.

WebGPU Computer Vision RunPod

BatemanAI

Built a financial AI assistant that acts as a personalized bank teller, with a generative UI that composes dashboards around what each user actually cares about — a college student and someone nearing retirement see entirely different views. Implemented a Twilio phone system for inbound and outbound calls: when the agent detects financial risk, like an upcoming rent payment colliding with excessive discretionary spending, it proactively calls the user with the issue and concrete steps to address it, and inbound callers can manage their finances entirely over the phone.

AI Agents Generative UI Twilio

Global Education Inequality Monitoring System

unesco-geims.org

Built a monitoring system that ingests millions of educational data points — faculty-to-student ratios, literacy rates, enrollment, internet access — for regions worldwide and distills them into an sEQI score (0–1) using an original equation, from extremely poor to ideal education quality. A dashboard heatmap shades each region by its score to help policymakers direct resources, with AI-generated insights suggesting concrete interventions. Pitched to UNESCO, earning an invite to Digital Learning Week 2025 in Paris, and presented on UNESCO's behalf at IEEE's WEEF & GEDC in South Korea.

Data Science sEQI UNESCO

Discordance

Built an MCP tool suite giving Owkin's K Pro AI Scientist platform a queryable knowledge graph for olfactory receptors in cancer, a domain with no existing infrastructure. Rather than collapsing conflicting evidence into a false-confidence answer, the system detects contradictions between primary sources, weights evidence by strength and replication count, and generates confidence-qualified rules instead of universal claims — pausing mid-session to ask the researcher via MCP elicitation when adjudication hits a genuine deadlock. Evidence was pulled and verified across primary literature, TCGA, PDB, ChEMBL, and patent data, with a post-run fabrication checker built after catching specific hallucinations in live testing. Covered three receptors (OR51E2, OR2H1, OR51B4), demoed with a contested-vs-consensus tension map, and evaluated at 100% on a 27-point rubric across six ground-truth queries.

MCP Knowledge Graphs Owkin