CARDIO4Cities / AI engineer case study

City research workspace

For: City Leads preparing government and healthcare meetings
In: a city name. Out: a cited brief, open questions, a knowledge graph, Q&A
Rule: every finding traces to a source; every unknown stays visible

8agents, 3 of them gates

3datastores: ledger, vectors, graph

4live search sources: Tavily web search + 3 scholarly APIs

$0models, search, storage, hosting on free tiers

Live research · open-weights models · no pre-seeded city data 01 / 08

The problem

Today

Days of reading portals, papers and policy
Ends as one person’s unverifiable document

The real risk: confidently wrong

A national figure quoted as a city one
A statistic nobody can source

“Understanding a city” = 5 areas

Area	Question
Burden	How common are CVD, hypertension, diabetes?
Programmes	What screening already runs?
Policy	Which NCD plans are in force?
Access	Facilities, workforce, pathways to care?
Actors	Which institutions matter?

02 / 08

AI solution: a team of agents

Planner→Discovery→Crawl gate→Reader→ Claim agent→Fact checker→Sufficiency→Brief

↺ Sufficiency sends the Planner back for missing areas · bounded by depth

Agent	Job	Hands on
Planner	Turns the city into searches per area	Search queries
Discovery	Tavily web search for government and programme pages; OpenAlex, Europe PMC, Wikipedia for studies. Ranks by authority	Candidate sources
gate Crawl gate	Reads robots.txt before any fetch	Permitted sources
Reader	Fetches HTML, XML, PDF	Clean text
Claim agent	Proposes findings, each with an exact quote	Candidate claims
gate Fact checker	A different model judges each claim	Published or withheld
gate Sufficiency	Are 3 of 5 areas covered?	Publish, or search again
Answer agent	Answers questions from stored evidence only	Cited answer, or “not enough evidence”

Stack: LangGraph · open-weights models on Groq · Tavily + scholarly APIs · SQLite, Qdrant, Graphiti + Neo4j Sandbox · FastAPI · Docker on Render. No agent both writes a claim and approves it.

03 / 08

Knowledge management and retrieval

Store	Holds	Answers
SQLite ledger	Runs, sources, crawl decisions, facts, gaps, metrics	“What did the system decide, and why?”
Qdrant	Verified evidence passages, filtered by city	“What evidence is near my question?”
Graphiti + Neo4j Sandbox	Organisations, programmes, policies, indicators, places	“Who is connected to what?”

Question→Vectors + graph→Rerank→Answer agent→Cited answer or decline

Vectors and graph are queried together at question time; citations resolve to stored records, never to model memory
Graphiti is bi-temporal, so re-researching a city keeps what was previously true
If vectors or graph are down, the ledger still answers

04 / 08

Trustworthiness and evidence

Guard	Rule	If it fails
Permission	robots.txt is read before any fetch	Not fetched, listed with reason
Grounding	The quote must be found in the source text	Claim discarded
Numbers	Every number in a claim must be in its quote	Unsupported
Independence	The checker never wrote the claim	Withheld → gap
Scope	National evidence is not a city figure	Published with a warning

Hyderabad · passed“55% had hypertension”22/40 and 55% both appear in the quote; marked partially supported because the sample is one hospital.

Hyderabad · withheld“Meetings held in Hyderabad”The checker read the quote: five different cities. Recorded as a gap.

05 / 08

Worked example and evaluation: Hyderabad

Recorded run · 13 Sep 2026 · Balanced depth · 1 round · 51 s · docs/examples/hyderabad-research-run.json

80%of claims survived checking

100%of findings are city-level

3 / 5areas covered; policy and access became gaps

4 / 4checked by a model, not rules

From 46 sources to 4 findings

Discovered46

Permitted45

Claims5

Published4

Withheld1

Finding · burden · partially supported“Common comorbidities were hypertension (22/40, 55%), dyslipidemia (19/40, 47.5%), and diabetes mellitus (9/40, 22.5%).”Nizam’s Institute of Medical Sciences (Cureus, 2026). Confidence 0.70: a hospital sample, not city prevalence.

Honest readingGrounded is not representative, and the brief says so. This run predates Tavily: all 46 sources were scholarly, which is why policy and access came back empty. Tavily now adds government and programme pages to every run.

06 / 08

User experience

A City Lead, not an engineer

1Type“Hyderabad, India”

2Pick depthQuick, Balanced, Thorough

3Watch10 steps in 4 plain phases

4CheckEach finding opens to its quote

5Ask & take awayCited Q&A, downloadable brief

Depth	Sources read	Rounds	Web searches	Time
Quick	5	1	3	42 s measured
Balanced	up to 8	up to 2	up to 10	51 s measured (Hyderabad, 1 round)
Thorough	up to 12	up to 2	up to 12	~4–8 min est.

Degraded modes are shown, not hidden: a banner appears when a model, store or search provider is unavailable.

07 / 08

Trade-offs and limitations

Decisions: trust is never traded; time and cost are

Decision	Effect
Rules before models	robots.txt, quote and number checks cost no tokens
Search APIs, not scraping	Tavily + scholarly APIs: sanctioned, no captchas
Rate limit ≠ outage	Fail over to a free model; extraction 383 s → seconds
Brief first, graph later	The user never waits for the graph

Limitations and next steps

Limitation	Next
Grounded ≠ relevant	Reranker, labelled city evaluation set
Web search returns news, commerce	Authority ranking and fact checker filter; tune domains
Free tiers: rate limits, sleep, Sandbox expiry	Paid tiers
SQLite, no auth, no human review	PostgreSQL, authentication, review queue

Demo path

Unseen city→Watch the agents→Finding → quote→Open questions→Graph→Ask→Download

08 / 08