A selection of production ML systems we have designed, trained and shipped alongside domain experts in healthcare, banking and operations.
Neck ROI + multi-color-space feature extraction
Each patient's neck skin patch is decomposed across 5 color spaces (RGB · CMYK · HSL · LAB · XYZ) yielding 15+ per-channel statistics — a rich, non-invasive substitute for a blood draw.
RGB-G channel vs HbA1c class (dark-skin cohort)
Clear separation between Normal (0), Pre-diabetic (1) and Diabetic (2) clusters proves the green channel alone carries measurable glycemic signal — the model isn't guessing.
Random Forest feature importance
Hue channels (N_HSV_H, N_HSL_H) dominate the top of the ranking — melanin-normalized hue is the single strongest predictor, which is why the model generalizes across skin tones.
Multi-site primary care network
Clinical Healthcare
Deployed neck-image texture analysis (GLCM + Random Forest) to flag pre-diabetic risk at intake, routing high-risk patients to HbA1c confirmatory testing.
RN / SW touch frequency vs. hospital admissions
Patients receiving more frequent nurse and social-worker touches show a visible downward slope in admissions — quantifying the ROI of proactive care-team engagement.
Population hospitalization distribution
~70% of the cohort sits at 0–3 admissions, but a long right tail drives most of the cost — the exact segment the rising-risk model targets for early intervention.
RPM device adherence by admission count
Adherence drops sharply as admission counts climb — falling RPM engagement is an early, actionable leading indicator of an impending hospitalization.
Metropolitan senior home-care provider
Clinical Healthcare
Analyzed RN/SW touch patterns, RPM adherence and hospitalization history for a senior home-care population. Models show that increased care-team interactions and higher device adherence correlate with fewer admissions, enabling proactive outreach before escalation.
Average paid amount by cost-group category
Adult-with-Disability members average ~$3,800/yr — the highest of any segment and 27% above general Adult spend. This is the cohort the risk model prioritizes for care management outreach.
Year-over-year hypertension charged amount ($M)
Total hypertension spend climbed from $1.40B (2020) to $1.52B (2021) — a $122M YoY increase that quantifies the ROI ceiling for a rising-risk intervention program.
Complexity of health visits — multi-level self-attention
Irregularly spaced visits mix facility, pharmacy and outpatient codes with rich semantic dependencies. The self-attention architecture learns cross-visit relationships to predict future disease and cost.
Regional health insurance provider (Denver metro)
Insurance & Payer Analytics
Built a multi-level self-attention model over longitudinal claims (diagnosis, procedure, medication codes + costs) to predict High-Need / High-Cost members and forecast next-year hypertension spend. Population analytics segmented Adult-with-Disability as the most susceptible cost group and surfaced 336 juvenile hypertensive cases for early intervention.
Academic medical center
Clinical Healthcare
Context-aware retrieval system grounded on internal protocols and PubMed-indexed research, answering complex clinical queries with cited sources.
Tertiary care hospital
Clinical Healthcare
Time-series transformer predicting sepsis and acute deterioration 6 hours ahead of bedside alarm criteria, integrated with the EHR alerting queue.
Health plan CAHPS analytics team
Clinical Healthcare
BERT pipeline that splits member reviews into medical vs. drug streams, tags sentiment, and classifies each comment into an area-of-concern (Customer Service, Getting Needed Care, Health Plan Rating, Rx access). Care coordinators see the dominant negative drivers per region and act on them.
Radiology informatics group
Clinical Healthcare
Multimodal LLM pipeline that ingests DICOM metadata (body part, weighting, plane) plus the image, and answers SME-authored structured questions about the study — grounded, non-speculative, and formatted for radiologist review.
Health plan care-management team
Clinical Healthcare
After predicting High-Cost High-Need members, unsupervised clustering groups them into actionable cohorts (e.g. hypertensive seniors in a specific ZIP band) with shared benefit plan, geography, chronic-condition and utilization patterns — so care teams intervene per cluster instead of per member.
Pharmacy analytics team
Clinical Healthcare
Predicts Proportion of Days Covered (PDC) adherence for chronic-medication members, excludes single-refill members per measure spec, and flags likely non-adherent patients weeks before the gap so outreach can be scheduled.
Primary care triage desk
Clinical Healthcare
Tesseract OCR reads handwritten prescriptions and clinical transcripts; a BERT medical-NER layer extracts drugs, symptoms and procedures; a classifier then recommends the correct downstream medical specialty department for the patient.
Senior home-care behavioral health team
Clinical Healthcare
Longitudinal analysis of social-worker interactions against pain, depression and loneliness scores; quantifies which patients experienced measurable relief and feeds a targeting model that prioritizes SW visits where impact is highest.
National retail bank
Banking & Finance
NER + Gen AI pipeline extracting entities from KYC packets, sanction lists and transaction narratives to enrich risk scoring workflows.
Digital payments platform
Banking & Finance
Graph neural network spotting anomalous transaction rings and account takeover patterns across millions of daily events.
Neo-lender
Banking & Finance
Gradient-boosted model combining bureau data, cash-flow signals and alternative features for fair, explainable credit decisions.
Automotive parts supplier
Cross-Industry
CNN/YOLO defect grading on stamped and assembled components, rejecting out-of-spec parts before they reach downstream assembly.
Fortune 500 logistics firm
Cross-Industry
Natural-language interface that translates business questions into validated SQL, executes on the warehouse and returns governed answers.
Global consumer brand
Cross-Industry
Chain-of-thought + RAG agent checking marketing claims against regional regulations and internal brand guidelines before publication.