Medical coding at claims scale
Healthcare AI · Production · ML engineer · medical-coding AI company · 2021–2025
Problem
Convert clinical documentation into the ICD and CPT codes that drive provider billing — without a human coder reading every case. Payers and hospital systems don't have the coder headcount to keep up with volume, and the long tail of specialty codes is where both accuracy and revenue live.
Build
PySpark pipelines on Databricks to move clinical cases from source systems through de-identification, feature prep, and labeling at scale. SageMaker services for prediction — model inference wrapped in the latency and reliability SLAs that a claims pipeline actually demands.
Result
10M+ cases moved through the pipeline in production. Real patients, real claims, real dollars on the other side of every inference.
What it taught me
In regulated industries, the model isn't the moat. Pipeline reliability is. The teams that won this problem weren't the ones with the best architecture — they were the ones whose data layer didn't rot.
Breast cancer detection in screening mammograms
Healthcare AI · Clinical deployment · ML engineer · radiology AI startup · 2019–2021
Problem
Surface suspicious regions in screening mammograms for a radiologist to attend to — not to replace the read, but to direct attention and catch the cases the eye tends to miss at the end of a shift.
Build
Faster-RCNN detector trained on a curated mammogram dataset. Integration directly into PACS workstations — the actual screen the radiologist reads from — rather than a separate portal no one would open.
Result
87% AUC on a thousand-mammogram test set. Deployed into clinical radiology workflow, not sitting in a demo environment.
What it taught me
Clinician trust is the deployment bottleneck, not accuracy. A model that's right 90% of the time but lives in the wrong tool gets ignored. A model that's right 80% of the time and renders inside the workflow gets used.
ML at Cellhub — AI inside a mobile carrier
Telecom AI · Current · ML engineer · Cellhub
What I do
Build ML systems at the intersection of AI and mobile networks. The patterns I shipped in healthcare — production pipelines, observable inference, models that survive contact with a regulated ops team — translate directly. The network just has more telemetry and fewer HIPAA rows.
What I can say publicly
Not much — specific systems stay inside. The Carrier Watch series is where I write about what's happening across the industry more broadly, and what I'm seeing is consistent with what I was shipping in healthcare a few years earlier.
Translating AI for the SMBs most vendors ignore
Small Business AI · Advisory · On the side · ongoing
Problem
Small and mid-size business owners are drowning in AI theater. Every vendor in their inbox claims "AI-powered." Most of what's being sold is a wrapper on a prompt, a workflow tool with ambitions, or a solution looking for a problem.
Approach
Sit in the business. Find the real workflow pain — usually not where the owner thinks it is. Then tell them, in plain English, which AI is worth a check and which isn't. Seven years of shipping production ML makes that call a lot easier to make honestly.
Representative situation
A family construction company: voicemails as the inbox, paper invoices as the system of record, the owner's head as the only source of truth. The AI that matters here isn't a chatbot. It's the automation that gets paper off the desk and voicemails into a searchable log — work that's been possible for a year and the industry is only now catching up to.
Want to talk about any of this?
If you're inside a carrier or mobility stack shipping ML, selling AI into a regulated industry, or running a small business trying to figure out what's worth buying — I want to hear from you.