Every system, with its architecture and its limits.
Grouped by the part of the pipeline each one is really about. Colour is consistent across the whole site: red is always a guardrail, green is always evaluation, amber is always a person in the loop.
- RetrievalGetting the right context in front of the model
- GenerationInference, synthesis and reasoning
- GuardrailValidation, safety checks and hard limits
- Human-in-the-loopWhere a person stays in the loop
- EvaluationMeasuring whether it actually works
Voice & clinical systems
2Post-Discharge Voice Agent
Calls patients 72 hours after discharge — and escalates on hardcoded clinical rules the AI is never allowed to touch.
100%
Sensitivity across 92 evaluation cases
Inbound Clinic Voice Agent
liveA real phone number that answers, books an appointment end to end, and hangs up.
88% → 100%
Task completion
Retrieval systems
5SEC Finance RAG
Citation-backed answers over 25,000 SEC filings, with the guardrails that make them trustworthy.
<500ms
Median query latency, cache miss
Vectorless RAG
Retrieval over SEC filings with no embeddings and no vector database at all.
top-K → 1–3 whole
Sections retrieved per query
Metadata-Aware RAG
Turning natural-language questions into database filters before search ever runs.
100%
Precision on filtered queries
RAG Document QA
The foundations — single and multi-document question answering with LangChain.
single + multi-document
Scope
MedSignal API
Hybrid retrieval across live FDA data, PubMed and a local index — hardened against adversarial input.
<11ms
Input validation
Agents & orchestration
3AnalystAI
Seven specialised agents that take a spreadsheet and a plain-English question to a boardroom answer.
7
Specialised agents in the pipeline
PharmAgent
A six-node agentic loop that grades its own retrievals and checks its own answers.
2–4 hours → <60s
Manual review time
26 Agents, 720 Simulated Hours
A multi-agent social simulation of how a layoff announcement actually propagates.
26
Stakeholder agents
Applied ML & evaluation
4Student Mental Health Early Warning
Flags struggling students three to five weeks earlier — and explains why for every single one.
0.975
XGBoost AUC
Customer Churn Prediction
Optimising for recall, because a missed churner costs more than a wasted retention offer.
0.96
Test accuracy
MNIST — 14 Networks
Fourteen architectures trained to find out what actually moves validation accuracy.
97.50%
Best validation accuracy
Health Insurance Premium Prediction
Regression on 1,338 records, where one binary feature dominates everything else.
~0.77
Linear regression R², test
Archive
Earlier work and smaller repositories. Kept public because the progression is part of the story.
| Repository | Language | Year | What it is |
|---|---|---|---|
| LinearRegression | Jupyter Notebook | 2026 | Forecasting mobile and tablet sales from advertising spend. |
| resumegenerator | Python | 2026 | Programmatic resume generation. |
| travel-wiki | Markdown | 2026 | A personal travel knowledge base maintained by an LLM. |
| Urban-Transit-Optimization | Jupyter Notebook | 2025 | PySpark over 10M+ NYC taxi records — trip duration, payment behaviour and congestion modelling. |
| NYSO-Electricity | Jupyter Notebook | 2025 | Time-series forecasting of NYISO electricity prices with PySpark and MLlib. |
| Loan-Performance-Data-Insights | Jupyter Notebook | 2025 | Exploratory analysis of loan performance data. |
| SEC-Corporate-Filings-Insights | Jupyter Notebook | 2025 | Early analytical work on SEC corporate filings — the starting point for the RAG systems above. |