Curious by default
I want to understand why a system behaves the way it does, not only whether a metric moved.
Welcome. I’m Nick.
Machine learning, NLP, and applied AI, from research to production systems.
About / Approach
I turn complex problems into useful systems.
Based in Chicago, I’m a machine-learning engineer and software builder interested in language, healthcare, and applied AI.
I work across the full path from messy data and experiments to reliable pipelines, APIs, and user-facing products.
What I enjoy most is taking an ambiguous problem and making it concrete: understanding what the data can support, choosing the right technical approach, and building something other people can inspect, use, and improve.
Currently focused on
Applied AI, data products, and intelligent software.
End to end
I’m comfortable moving from data and models to backend services, databases, deployment, and product interfaces.
Make it dependable
Clear assumptions, reproducible work, thoughtful evaluation, and maintainable software matter as much as the first demo.
Selected work
Products, models, and useful systems.
Product-focused engineering in progress, followed by selected research and published work.
Product thesis / Working title: SignalBrief
A morning decision layer, not another inbox.
The first use case is job-search inbox intelligence: surface recruiter outreach, interviews, assessments, deadlines, application changes, and follow-up opportunities without asking the user to scan another chronological feed. The same architecture can later support university, work, and personal mail.
01What requires attention today?
02What changed since yesterday?
03What important older item is still unresolved?
01 / Implemented foundation
A complete path from versioned inbox data to an evaluated, persisted, and reconstructable morning brief.
01 / Validate
Typed domain contracts
Pydantic separates provider facts, extractor claims, user state, and historical briefing context, and rejects invalid or unsupported states.
02 / Select
Point-in-time candidates
Hard gates remove resolved, dismissed, archived, or not-yet-arrived messages before explicit inclusion reasons are assigned.
03 / Rank
Deterministic baseline
Priority, deadlines, recency, and stable tie-breaking create a transparent benchmark for future learned rankers.
04 / Brief
Product-facing output
The pipeline assembles attention, update, and activity sections and renders the same typed result as human-readable text or JSON.
05 / Persist + evaluate
Versioned evidence
Chronological groups produce ranking metrics while PostgreSQL preserves exact briefing lineage for later historical reconstruction.
02 / Technical thesis
The differentiator will be the full decision lifecycle, but it starts with a defensible baseline.
Personalized learning to rank / next
Learn the order of each morning list rather than classifying emails independently, then compare it with the deterministic baseline and a zero-shot LLM reference.
Evidence-grounded extraction / planned
Future action and deadline claims must point back to supporting text, with abstention when evidence is weak instead of invented certainty.
Delayed-feedback learning / planned
Future feedback will separate immediate interactions from later outcomes while preserving point-in-time features and avoiding training-serving leakage.
Decision lineage / implemented
Immutable briefing runs and items retain the exact snapshot, analysis version, ranking decision, and workflow values required to reconstruct historical output.
03 / Modeling and evaluation
The evaluation machinery exists before the learned model does.
Model ladder
- 00Implemented: deterministic, explainable rule baseline
- 01Next: transparent logistic or ordinal priority model
- 02Then: gradient-boosted learning-to-rank model
- 03Ablation: structured features plus semantic embeddings
- 04Later: per-user calibration or contextual adaptation
- REFReference: zero-shot LLM ranking for quality, latency, and cost
Evaluation contract
- NDCG@5Implemented for ranked morning lists
- Urgent Recall@5Implemented safety metric for top-ranked urgent items
- MRRPlanned for the learned-ranking comparison
- CalibrationPlanned with probabilistic models
- Support ratePlanned for grounded extraction
- Latency + costPlanned for backend and model serving
04 / Reliability foundation
Preserve what the system knew and what it decided.
The persistence layer now stores immutable briefing runs and ordered items linked to exact message snapshots and analysis versions. A historical brief can be reconstructed from PostgreSQL without reading future mailbox or workflow state; full champion-versus-candidate replay remains a later MLOps stage.
01
Persist an immutable briefing run and its ordered decisions.
02
Link every item to the exact message snapshot and analysis version used.
03
Copy the mutable workflow values visible at decision time.
04
Reconstruct the same typed morning brief from PostgreSQL.
05
Next: re-run champion and candidate models on the preserved context.
06
Later: compare, promote, reject, or roll back with documented evidence.
05 / Privacy and boundaries
Email is sensitive; data minimization is part of the system design.
Required controls
- Delegated, read-only Microsoft Graph access with no Mail.Send permission.
- Encrypted secrets and storage; no OAuth tokens or raw message content in logs.
- Short, documented retention for raw bodies and a complete disconnect-and-delete flow.
- Public synthetic mailbox, fixtures, schemas, and aggregate metrics, never private emails or recoverable identifiers.
- Structured extraction boundaries to reduce prompt-injection risk from untrusted message content.
Intentionally outside version one
- Automatic replies, sending, deleting, or autonomous mailbox actions.
- Gmail and multi-provider support before the Outlook vertical slice is dependable.
- A generic conversational assistant unrelated to the morning briefing.
- Kubernetes, streaming infrastructure, or a feature store without demonstrated need.
- Claims that every important message will be detected; the original inbox always remains accessible.
06 / Delivery roadmap
Build one credible vertical slice, then add learning and operational depth.
Completed / Phase 1
Typed contracts, deterministic selection and ranking, a morning-brief CLI, and chronological ranking evaluation.
Completed / Persistence
Dockerized PostgreSQL, schema migrations, repository boundaries, versioned feedback, briefing lineage, historical reconstruction, and 32 passing tests.
Next / API layer
FastAPI transport, request-scoped transactions, historical briefing retrieval, priority correction, and structured logging.
Release 01–02
Read-only Outlook synchronization, private chronological labels, learned ranking, untouched evaluation, and a model card.
Release 03–05
Grounded extraction, native iOS experience, full model replay, monitoring, controlled release, and rollback.
Current status / Persistence foundation complete
The deterministic vertical slice now runs through a Dockerized PostgreSQL persistence layer with schema migrations, repository and service boundaries, append-only feedback history, immutable briefing lineage, and verified historical reconstruction. FastAPI, Outlook synchronization, learned ranking, grounded extraction, cloud deployment, monitoring, and the iOS client remain next.
Project thesis / DivvyAI
Record first. Predict later.
Divvy’s public feeds only describe the present, so the history needed to train a forecasting model doesn’t exist yet. DivvyAI records it: both station feeds, every 15 minutes, with the original responses preserved and loaded into BigQuery.
01How many bikes will this station have in 30 minutes?
02Which stations are about to run empty?
03How do availability patterns change by hour, day, and location?
01 / How it works
One scheduled job: capture, archive, load, verify.
01 / Schedule
Every 15 minutes
Cloud Scheduler triggers a containerized Python job on Cloud Run.
02 / Capture
Both station feeds
Station information and status, with source URLs and UTC receipt times.
03 / Archive
Original bytes first
One archive per capture in Cloud Storage, with a manifest and a stable capture ID.
04 / Load
Batch into BigQuery
Append-only loads into an ingestion-time partitioned table.
05 / Verify
SQL validation
Capture-scoped queries confirm the warehouse matches what was received.
02 / Engineering decisions
Built for trustworthy, replayable training data.
Data integrity
- Original responses are archived before parsing, so any capture can be replayed.
- Source, receipt, and warehouse times are kept separate to avoid training leakage.
- 164 automated tests on synthetic and recorded data; Ruff and mypy enforced.
Security and cost
- Keyless authentication through a dedicated, least-privilege service account.
- Batch loads instead of streaming inserts, with partition filters required on every query.
- Billing budget and alerts configured before any resource was created.
03 / Roadmap
From reliable data collection to availability forecasting.
Completed / Pipeline
Tested collector and an end-to-end Cloud Run pipeline into Cloud Storage and BigQuery.
Live / Scheduling
Automatic captures every 15 minutes.
Next / Recovery
Archive replay, load reconciliation, and a retention and cost review.
Then / Features
dbt models that turn raw snapshots into tested, time-aware features.
Goal / ML forecasting
A model predicting station availability 30 minutes ahead, evaluated against persistence and historical baselines.
Current status / Collecting history
Captures run every 15 minutes, building the dataset the forecasting model will be trained on. Independent project using Divvy’s public data feeds; not affiliated with Divvy or Lyft.
Research
Research with real outcomes.
Multimodal clinical AI, published predictive modeling, and explainable NLP for healthcare policy research.
Simplified system architecture
Speech audio
Wav2Vec2 representations
Transcript
LLaMA text embeddings
Soft-prompt adapter
Cross-modal projection and trainable fusion
MCI classifier
Lightweight prediction layers
Low-resource clinical learning
Explore whether complementary speech and language signals can support earlier MCI detection when clinical datasets are small and imbalanced.
Parameter-efficient fusion
Keep the speech and language backbones frozen while training cross-modal projections, soft-prompt adapters, and lightweight classification layers.
Reproducible clinical experiments
Compare model scale, task type, transfer strategy, and imbalance handling across PROCESS-2 and ADReSS picture-description and verbal-fluency tasks.
Status: ongoing research. Performance findings are intentionally omitted until the experiments are validated and ready to share.
Final cross-validated performance
2,013
Patient records
.86
Accuracy
.90
Recall
.87
F1 score
.94
AUC
DLNN
.86
Random forest
.74
Decision tree
.65
End-to-end model development
Built the preprocessing, feature-engineering, imbalance-handling, model-validation, and benchmark workflows used to develop and assess the neural network.
Regularized clinical DLNN
Combined ReLU hidden layers, L2 regularization, batch normalization, sigmoid output, binary cross-entropy, early stopping, and model checkpointing.
Designed for imbalanced data
Addressed an 85/15 class split with SMOTE, evaluated with five-fold cross-validation, and benchmarked against random-forest and decision-tree baselines.
Publication outcome · Open access
Deep learning neural network prediction of postoperative complications in laparoscopic right hemicolectomy
Read publication ↗
Two-stage prediction and explanation system
Scenario + profile
Duration, relative reimbursement, professional role, and degree
Respondent evidence
Original open-ended responses, human-coded themes, and sentence embeddings
Predictive core
Structured benchmark → embedding ablation → nested threshold selection
Willingness estimate
Out-of-fold prediction with each respondent kept in one group
Grounded LLaMA layer
Blind JSONL generation → grounded rationale → separate error audit
Current respondent-grouped evaluation
214
Decisions
72
Respondents
.692
Macro-F1
.579
Minority recall
Nested threshold tuning improved macro-F1 from 0.659 to 0.692 and minority-class recall from 0.386 to 0.579, with every respondent kept entirely within a single fold.
From survey export to an auditable data system
Converted the survey export into linked scenario- and participant-level datasets covering 214 reimbursement decisions from 72 respondents. The pipeline preserves repeated decisions, original open-ended responses, codebook context, and model outputs without inventing outcomes for incomplete respondents.
Test each source of signal separately
Completed a controlled ablation across reimbursement representations, respondent characteristics, human-coded themes, and SentenceTransformer embeddings. Fold-isolated dimensionality reduction supports the text features, while a transparent structured model remains the benchmark for the selected text-augmented candidate.
Prevent leakage, and tune decisions honestly
Implemented nested respondent-grouped cross-validation so every decision from the same person stays in one fold. Preprocessing, PCA, and threshold selection are fitted only inside training data before producing out-of-fold predictions and targeted error cases.
Current system state
A completed predictive pipeline first. A separate, outcome-blind explanation workflow second.
- 01Built linked datasets for 214 decisions from 72 respondents
- 02Completed structured, human-coded, and semantic-feature ablations
- 03Implemented nested grouped validation with fold-isolated preprocessing, PCA, and threshold tuning
- 04Frozen out-of-fold predictions and scenario-level evidence records for explanation
- 05Built a two-stage LLaMA workflow with blind JSONL generation, structured outputs, and separate error audits
- 06Next: expert review, targeted error analysis, and full-sample explanation validation
Status: active research and validation. Predictive results are from respondent-grouped nested cross-validation; expert review and full-sample validation of the generated explanations remain ongoing. Predictions and rationales describe model-supported associations, not causal reasons, clinical recommendations, or deployment readiness.
Get in touch