Share
Embed
Slide 1
Slide 1 text
🧑⚕ Clinics on Rails
Alex Watt
@alexcwatt
Slide 2
Slide 2 text
No content
Slide 3
Slide 3 text
No content
Slide 4
Slide 4 text
No content
Slide 5
Slide 5 text
No content
Slide 6
Slide 6 text
No content
Slide 7
Slide 7 text
🎓
2,200 families self-supporting
Slide 8
Slide 8 text
No content
Slide 9
Slide 9 text
2025 Pittsburgh Team
Slide 10
Slide 10 text
No content
Slide 11
Slide 11 text
Benefits of clinical data
● Set volunteer expectations
● Plan for pharmacy
● Understand community health over time
Slide 12
Slide 12 text
Requirements doc
January 2024
Slide 13
Slide 13 text
No content
Slide 14
Slide 14 text
Surveyed options
Throughout 2024
Slide 15
Slide 15 text
BackpackEMR
✅ Designed for mobile medical clinics
❌ PII stored outside of Ethiopia, could not host in country, limited reporting
Slide 16
Slide 16 text
Google Sheet
✅ Free, ubiquitous, simple
❌ Relational data, limited data validation, more difficult integrations, permissions
Slide 17
Slide 17 text
OpenMRS & others
✅ Powerful open source package
❌ Seemed non-trivial to deploy and customize
Slide 18
Slide 18 text
It had to be simple.
Slide 19
Slide 19 text
No content
Slide 20
Slide 20 text
No content
Slide 21
Slide 21 text
Built in a week (January 2025)
Monday rails new
Tuesday Finished data models
Wednesday Deployed with Hatchbox
Thursday User policies & launch
Friday Addressed feedback
Slide 22
Slide 22 text
No content
Slide 23
Slide 23 text
Project Impact
● Volunteers can help with medical data
● Medical teams see their impact
● Future teams are better prepared
Slide 24
Slide 24 text
1,583 patient encounters documented
in 5 communities in Addis
Slide 25
Slide 25 text
Innovation Tokens
acw.sh/boring
Slide 26
Slide 26 text
I spent only one token — on search.
Slide 27
Slide 27 text
Neoplasms
Diseases of the blood or blood-forming organs
Endocrine, nutritional, or metabolic diseases
etc.
Slide 28
Slide 28 text
No content
Slide 29
Slide 29 text
Simple Hybrid Search
+
semantic fuzzy match
Slide 30
Slide 30 text
Simple Hybrid Search
+
semantic fuzzy match
Query: fizzy
sparkling water
soda
frizzy hair
Fizzy’s Diner
Slide 31
Slide 31 text
Simple Hybrid Search
+
informers gem
(sentence-transformers)
Jaro-Winkler
Slide 32
Slide 32 text
sentence-transformers/all-MiniLM-L6-v2
91 MB
1 vCPU and 1 GB RAM
Slide 33
Slide 33 text
Embedding computation
MODEL_NAME = "sentence-transformers/all-MiniLM-L6-v2"
EMBEDDER = Informers.pipeline("embedding", MODEL_NAME)
def embedding(text)
embeddings_array = EMBEDDER.(text)
Vector[*embeddings_array]
end
Slide 34
Slide 34 text
Embedding computation
MODEL_NAME = "sentence-transformers/all-MiniLM-L6-v2"
EMBEDDER = Informers.pipeline("embedding", MODEL_NAME)
def embedding(text)
embeddings_array = EMBEDDER.(text)
Vector[*embeddings_array]
end
Slide 35
Slide 35 text
Embeddings for diagnoses
diagnoses_with_embeddings = Diagnosis.all.map do |diagnosis|
text = "#{diagnosis.name} (e.g., #{diagnosis.examples})"
embedding = embedding(text)
{ diagnosis: diagnosis, embedding: embedding }
end
Slide 36
Slide 36 text
Query embedding
QUERY_PREFIX = "Represent this sentence for searching
relevant passages: "
query_embedding = embedding(QUERY_PREFIX + query)
Slide 37
Slide 37 text
Search with embeddings and fuzzy text search
LIMIT = 3
diagnoses_with_embeddings.filter_map do |d|
semantic = cosine_similarity(query_embedding, d[:embedding])
fuzzy = fuzzy_score(query, d)
semantic_relevant = semantic >= MIN_SEMANTIC_SCORE
fuzzy_relevant = fuzzy >= MIN_FUZZY_SCORE
next unless semantic_relevant || fuzzy_relevant
[d[:diagnosis], [semantic, fuzzy].max]
end.sort_by { |_, score| -score }.first(LIMIT).map(&:first)
Slide 38
Slide 38 text
Search with embeddings and fuzzy text search
LIMIT = 3
diagnoses_with_embeddings.filter_map do |d|
semantic = cosine_similarity(query_embedding, d[:embedding])
fuzzy = fuzzy_score(query, d)
semantic_relevant = semantic >= MIN_SEMANTIC_SCORE
fuzzy_relevant = fuzzy >= MIN_FUZZY_SCORE
next unless semantic_relevant || fuzzy_relevant
[d[:diagnosis], [semantic, fuzzy].max]
end.sort_by { |_, score| -score }.first(LIMIT).map(&:first)
Slide 39
Slide 39 text
Search with embeddings and fuzzy text search
LIMIT = 3
diagnoses_with_embeddings.filter_map do |d|
semantic = cosine_similarity(query_embedding, d[:embedding])
fuzzy = fuzzy_score(query, d)
semantic_relevant = semantic >= MIN_SEMANTIC_SCORE
fuzzy_relevant = fuzzy >= MIN_FUZZY_SCORE
next unless semantic_relevant || fuzzy_relevant
[d[:diagnosis], [semantic, fuzzy].max]
end.sort_by { |_, score| -score }.first(LIMIT).map(&:first)
Slide 40
Slide 40 text
No content
Slide 41
Slide 41 text
No content
Slide 42
Slide 42 text
What I hope you take away
Slide 43
Slide 43 text
#1: Your coffee ritual is (probably) weak
☕
Slide 44
Slide 44 text
#2: End-to-end ASAP
📅
Slide 45
Slide 45 text
#3: Simple software is great
👨🔧
Slide 46
Slide 46 text
#4: Open-weight models are awesome
✨
Slide 47
Slide 47 text
#5: You can just do things
💪
Slide 48
Slide 48 text
አመሰግናለሁ
acw.sh/act