HealthSurface: Health news, funding and jobs, labeled by source type.

What It Does

One reading list with three tabs that share the same Sector and Focus area tags. There are no accounts, no personal data and no medical advice, and every item links to its original source.

• News: 160 stories from openFDA, medRxiv, PubMed, The Conversation, and the CMS and CDC newsrooms. Every story carries a Source type label (Press release, Preprint, Peer-reviewed study, Regulatory action or Reported news), with a tooltip explaining each one. A daily "In focus" brief picks the top stories.
• Funding: 174 records from SEC Form D filings and NIH small business research grants. Amounts come from the filings themselves, and a round or investor is shown only when a source states it.
• Jobs: 1,460 open roles at 69 health companies from verified company job boards (Greenhouse, Lever, Ashby) and Remote OK, filterable by function, 36 job types, employment type, country and remote.
• Daily summary you can listen to: a short AI-written summary of the day's top stories and funding, linking every source it mentions and read aloud by an Amazon Polly voice.
• Ask what's new: a chat on every tab that answers only from the stored stories, filings and job posts, lists its sources with their Source type, and declines "should I" questions. Questions are never stored or logged.

Designing for Trust

Language models are good at reading and bad at being trusted, so I split the work on purpose. Code decides the facts: the Source type label (from where an item came), links, dates, deduplication, the free-to-read check and job board verification. Amazon Nova handles the reading: sector and focus areas, a two-sentence summary, whether a headline announces funding, and job function, type and seniority. Every model reply is checked against the taxonomy, and off-list answers are dropped instead of trusted.

Funding facts are checked in code. Early on, the model labeled a $1B Sanofi and Regeneron partnership as a "Series B" round. Now a story only counts as funding if the text says raised, closed or secured, and the amount, round and investors must each appear in the source text or they're left blank. The model can't invent a number.

How I Used AWS

HealthSurface runs entirely on AWS, defined as code in a single AWS SAM template:

• Amazon Bedrock (Nova Lite): classifies every story and job (sector, focus areas, job function, type and seniority), writes the two-sentence summaries, flags funding headlines, writes the daily summary, and powers the Ask chat. It costs about $0.0001 per story, so a full refresh costs about a cent.
• Amazon Bedrock Guardrails: checks every model output for medical advice and dosing, with a separate guardrail on the Ask chat. Guardrails are deliberately not applied to input, because real headlines mention doses and treatments and an input filter would block real news.
• Amazon Polly: reads the daily summary aloud behind the Listen button.
• AWS Lambda (Python 3.14, arm64): three functions: an ingest function that pulls from the news sources, SEC EDGAR, NIH RePORTER and job boards; a read-only API; and the Ask function. arm64 is the cheapest compute option.
• Amazon EventBridge Scheduler: runs ingestion twice a day, with no retries.
• Amazon DynamoDB (on-demand): stores news, funding, jobs and run metadata, holds a lock that prevents overlapping ingest runs, and enforces rate limits on the Ask chat.
• Amazon S3: hosts the static site behind Origin Access Control and stores the daily summary audio.
• Amazon CloudFront: serves the site and routes /api/* to API Gateway, so there's one public URL and no CORS. Deploys finish with a CloudFront invalidation.
• Amazon API Gateway (HTTP API): fronts the API and Ask Lambdas, cached for 5 minutes and throttled.
• AWS SAM: all infrastructure as code in one template, shipped with sam deploy.
• AWS Budgets: a $25 budget alert, alongside hard caps of 100 stories and 200 jobs per run, so only new items are ever classified.
• AWS sign-in (aws login): how the coding agent connected to the account to run every deploy.

Built with a Coding Agent

I wrote the product brief and shipped HealthSurface in one evening with Claude Code (Claude Opus 5.5), which ran every AWS step itself: the identity check, the Bedrock model check, the budget, the guardrail, every sam deploy, the S3 uploads and the CloudFront invalidations. Each build step was tracked as a Linear issue through the Linear MCP connector and deployed on completion, so there was always a working live version. The first AWS command ran at 4:42 PM; all three tabs were live with real data by 5:16 PM.

The rest of the evening went to quality, including problems a tutorial wouldn't cover:

• A new-account Lambda quota: the first deploy rolled back because new accounts must keep 10 concurrency units free. Reserved concurrency was replaced with a DynamoDB run lock.
• Bedrock account verification: Nova showed as active but the first real call was denied, so a keyword fallback classifier kept the site working until the model came online. Today 165 of 169 stories and 97% of jobs are model-classified.
• Dates from the future: PubMed's sort date is the journal issue date, sometimes months ahead, so ingestion switched to the indexed date.
• Source terms: I had every source's terms reviewed, dropped a feed that bans automated access, and added the credits NCBI and Remote OK require.

Tools Used

Python, Amazon Bedrock (Nova Lite), Bedrock Guardrails, Amazon Polly, AWS Lambda, EventBridge Scheduler, DynamoDB, S3, CloudFront, API Gateway, AWS SAM, AWS Budgets, Claude Code, Linear MCP, GitHub. Backed by 53 unit tests.

What's Next

An MCP server and an Alexa+ daily voice briefing that reuse the classifier as-is; it's a plain Python module with no AWS code, and the model is passed in as a function. Also on the roadmap: ClinicalTrials.gov and Europe PMC as sources, email digests by focus area, and paid alerts for investors and operators while the reading list stays free.