Research assistant, KI· 2025
AI reads the dosing notes. Code checks every row.
Free-text Swedish dosage instructions, turned into structured records a research group could use. The model reads; code checks every row it returns.
The outcome
Every one of 1,000 prescription records came back: 99.4% as valid, rule-checked output and 1.1% flagged for a person to review. The old process had dropped about 90% of them.
- 99.4%came back as valid, rule-checked output
- 1.1%flagged for a person to review
- 0records lost
- ~90%lost by the old process
- Role
- Research Assistant, Applied AI for Medical Data
- When
- Apr–Oct 2025
- With
- Zheng Chang's group, psychiatric pharmacoepidemiology · built solo
- Proves
- Builds whatever it takes
Layers I owned
- Discovery
- Strategy
- Architecture
- Build
- Rollout
01Why I built it
The group studies how psychiatric medicines are used, and prescriptions are the raw material. Their old pipeline couldn't read free-text dosage instructions, so most of a Swedish prescription dataset never reached the research.
02How I did it
I agreed labelling rules with the researchers, wrote a JSON schema for a structured prescription, and iterated prompts on Gemini 2.5 Pro. Every row the model returned then went through a deterministic layer: clinical logic rules, a duplicate check, a confidence score and review flags.
- Free textSwedish dosage
- LLM parsingGemini 2.5 Pro
- Rules checklogic · duplicates
- Structured recordor 1.1% to review
ObservationThe model was strongest exactly where rules break: reading Swedish phrasing, working out titration schedules and normalising indications to English.
03What got hard
Titration schedules hide arithmetic, like doses that step up week by week.
So I let the model work the schedule out, then checked the result against clinical logic rules in code.
Some regimens were unusually high.
So I flagged them for clinical review instead of letting the pipeline decide.
04What I didn't choose
Keep a rules-only parser
That was the old process, and it lost about 90% of the records.
Trust the model's output
Integrity and safety stay in code. Every row is checked, and doubtful ones go to a person.
05Shipping it
I set the pipeline up for the group and ran it on their real prescription data. It's a research pipeline, not a production rollout.
What this doesn't prove
- 99.4% is the share of valid, rule-checked output, not accuracy against a gold standard.
- A validated research pipeline, not a live clinical system.
- The data isn't public, so there are no screenshots of real records.
- Gemini 2.5 Pro
- JSON Schema
- Rule-based validation
- Hash de-duplication
- Confidence scoring