Asha — Medical Intelligence by DNAi Systems

Evidence-grounded medical AI, accessible via the A2A protocol.

Asha is a fiduciary medical AI agent backed by ~87M medical-domain knowledge vectors — a curated slice of the larger Citadel corpus (121M+ vectors across ~591 collections covering medicine, pharmacology, research literature, clinical guidelines, and structured codings). Every response includes structured provenance — source collections, evidence count, and predicate classification — so you can verify what grounded the answer.

Built by physician co-founders. Patent allowed: US 19/290,471.

Agent Card

Live at:

https://api.askasha.org/.well-known/agent-card.json?agent_id=asha

Fleet discovery (all 11 public agents in one card):

https://api.askasha.org/.well-known/agent-card.json

A copy of Asha's card is included in this repo at agent-card.json.

Skills

Skill	ID	Description
Medical Q&A	medical-qa	Differential diagnosis, lab interpretation, mechanism of action, comorbidity management
Drug Interaction Check	drug-interaction-check	Polypharmacy safety evaluation using DailyMed and FDA drug label data
Clinical Guideline Lookup	clinical-guidelines	ADA, USPSTF, AHA, WHO guidelines with citation support
Evidence Synthesis	evidence-synthesis	Multi-source literature synthesis across PubMed, OpenAlex, StatPearls

Quick Start

1. Get an API Key

curl -X POST https://api.askasha.org/api/a2a/signup \
  -H "Content-Type: application/json" \
  -d '{"email": "you@example.com", "name": "Your Name", "tier": "free"}'

The response includes your API key (shown once — save it).

2. Send a Query

curl -X POST https://api.askasha.org/a2a/v1/message:send \
  -H "Authorization: Bearer YOUR_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "message": {
      "role": "user",
      "parts": [{"text": "What is the mechanism of action of metformin?"}]
    },
    "metadata": {"agent_id": "asha"}
  }'

3. Parse the Response

Every response returns a Task with two artifacts:

Response artifact — the clinical answer with citations.

Provenance artifact — structured metadata:

{
  "sources": ["pubmed_abstracts", "dailymed_drug_labels", "clinical_guidelines"],
  "evidence_count": 48,
  "contract_hash": "...",
  "predicate_scope": ["medical_general"]
}

See examples/python_client.py for a complete Python example.

A2A Protocol Compliance

Requirement	Status
Agent Card at /.well-known/agent-card.json	Served
POST /a2a/v1/message:send	Implemented
GET /a2a/v1/tasks/{id}	Implemented
GET /a2a/v1/tasks (list, paginated, scoped to caller)	Implemented
POST /a2a/v1/tasks/{id}:cancel	Implemented
GET /a2a/v1/health (503 when unhealthy)	Implemented
Bearer auth with securitySchemes	Implemented
A2A-Version: 1.0 response header	Implemented
Task state lifecycle	SUBMITTED, WORKING, COMPLETED, FAILED, CANCELED
Task scoping to caller	Implemented (v1.0 §4.3)

Evaluating Asha against your current medical AI

Already on UpToDate, OpenEvidence, Glass, an OpenAI-compatible internal stack, or anything else? Here is a 30-minute, no-credit-card, side-by-side test.

What makes this different from a typical LLM bake-off

Asha returns structured provenance with every response. That means you can compare not just the prose, but the actual epistemic substrate behind it. Most general-purpose APIs do not let you see this.

Dimension	Asha returns	Typical chat API returns
Source collections that grounded the answer	Yes (provenance.sources)	No
Count of evidence items used	Yes (provenance.evidence_count)	No
Predicate scope (medical_general, drug, guideline, etc.)	Yes (provenance.predicate_scope)	No
Verifiable fiduciary contract hash	Yes (provenance.contract_hash)	No
Independent claim falsification	Yes (POST /api/feng/falsify, see feng-a2a)	No

The 5-step protocol

1. Get a free key (no credit card, 50 queries / month).

   curl -X POST https://api.askasha.org/api/a2a/signup \
     -H "Content-Type: application/json" \
     -d '{"email":"you@example.com","name":"Your Name","tier":"free"}'

2. Pick a representative evaluation set. 20–50 questions from your real query log. Mix easy (drug MoA, guideline lookup), medium (clinical guidelines, contraindication chains), and hard (multi-system differentials, edge-case interactions). Include questions you have seen your current provider hallucinate on. A starter is at examples/eval_set.txt.

3. Run the parallel harness at examples/evaluation.py. Edit query_other_api() to match your current provider's auth + body format (most are OpenAI-compatible chat completions; if so, set OTHER_API_BASE and OTHER_API_KEY and you're done).

   ASHA_API_KEY=ak_...        \
   OTHER_API_BASE=https://your-current-api  \
   OTHER_API_KEY=...          \
   EVAL_SET=examples/eval_set.txt \
     python examples/evaluation.py

   Output: ab_results_<timestamp>.json with both providers side-by-side per question, plus latency p50/p95 and Asha evidence-count distribution.

4. Score the dimensions that matter. Most teams care about some subset of:

   Dimension	How to score
   Factual accuracy	Manual review against an authoritative reference (UpToDate, primary literature)
   Citation quality	Asha returns sources and evidence_count directly. For other providers, parse cited references and verify they exist (some APIs hallucinate citations).
   Hallucination rate	Per response, count fabricated drug names, dosing claims, made-up guidelines. Asha's pre-generation suppression + post-generation verification is designed to drive this to zero on grounded queries.
   Latency p50 / p95	Auto-collected by the harness
   Cost at expected volume	Asha is flat-rate metered ($0 / $49 / $199 / Custom — see Pricing). Compare against per-call or per-token billing on your current provider.
   Auditability	Can you re-derive what knowledge grounded the answer, weeks later, after the model has updated? Asha: yes (contract hash + sources are stable per response). Most: no.

5. Stress-test any specific claim with FENG. If either provider says something high-stakes (e.g. "Drug X is contraindicated with Drug Y"), run that exact claim through the FENG endpoint. You'll get a verdict — FALSIFIED, WEAKENED, CONDITIONAL, or UNFALSIFIED — with E-value, evidence-for, evidence-against, and the collections searched. This is the deepest single audit step available across any medical AI vendor today.

What we tell evaluators honestly

• Asha is strong on: drug pharmacology, drug-drug interactions, clinical guidelines, evidence synthesis, anything where citation fidelity matters more than chatty bedside manner.
• Asha is intentionally conservative on: definitive diagnosis, medication dosing, anything that requires a licensed clinician's judgment. The fiduciary medical contract refuses to fabricate any of those rather than guess. If your current provider gives confident definitive diagnoses on demand, that is a difference, not a defect — measure hallucination rate before treating it as a feature.
• Latency: retrieval-augmented generation across 87M+ medical vectors is not free. Expect a few seconds end-to-end. The harness will surface exact p50/p95 numbers for your set so you can weigh the trade-off explicitly.

Migration path if you decide to switch

• The A2A surface is bearer-auth + JSON. Most teams swap the URL and Authorization header and ship.
• Run a 2-week shadow window: route real production queries to both, log both responses, review divergences. The harness's output JSON is shaped for exactly this.
• Keep your current API as a fallback; rotate Asha keys with POST /api/a2a/rotate-key if a key is ever leaked.
• For high-volume production deployments, contact us via dnai.systems for an Enterprise tier with custom limits and SLAs.

Connect to Google Gemini Enterprise

Asha (and any of the 11 DNAi agents) registers as a "Custom agent via A2A" inside a Google Gemini Enterprise app. Once registered, the agent shows up to your end users in the Gemini Enterprise web app and Gemini composer surfaces.

The agent card in this repo (agent-card.json) is shaped to be maximally compatible with the Gemini Enterprise registration parser: it carries both the A2A v1.0 supportedInterfaces[] shape AND the flatter top-level url + protocolVersion fields used by the Gemini Enterprise console example, plus iconUrl and documentationUrl.

Console (recommended)

1. In the Google Cloud console, open Gemini Enterprise → click your app → Agents → Add Agents.
2. In Choose an agent type, pick Add → Custom agent via A2A.
3. Paste the contents of agent-card.json into the Agent card JSON field.
4. Click Preview agent details → Next.
5. If your users will need Asha to access Google Cloud resources on their behalf (Drive, Docs, BigQuery, etc.), add the OAuth client credentials. Otherwise click Skip & Finish. Asha itself does not require Google OAuth — it authenticates with its own DNAi-issued API key (Bearer auth, declared in securitySchemes).

REST

If you'd rather automate it:

PROJECT_ID=your-gcp-project
LOCATION=global       # or us, eu
APP_ID=your-gemini-enterprise-app-id
ENDPOINT=${LOCATION}-discoveryengine.googleapis.com

curl -X POST \
  -H "Authorization: Bearer $(gcloud auth print-access-token)" \
  -H "Content-Type: application/json" \
  "https://${ENDPOINT}/v1alpha/projects/${PROJECT_ID}/locations/${LOCATION}/collections/default_collection/engines/${APP_ID}/assistants/default_assistant/agents" \
  -d @- <<EOF
{
  "name": "asha-dnai",
  "displayName": "Asha (DNAi)",
  "description": "Fiduciary medical intelligence with verifiable provenance.",
  "a2aAgentDefinition": {
    "jsonAgentCard": $(cat agent-card.json | jq -c . | jq -R .)
  }
}
EOF

That single command embeds Asha's agent card into the Gemini Enterprise registration request. Replace asha in the agent-card URL / agent_id metadata to register any of the other DNAi agents (Harley, Artha, Sage, Polymath, Lyra, Leo, Mira, Ren, Arohi, Ray).

What Gemini Enterprise users will see

• Asha appears in their Gemini Enterprise app's agent list with the icon at https://dnai.systems/asha-logo.svg.
• Queries route through Gemini's UI to https://api.askasha.org/a2a/v1/message:send over the A2A v1.0 protocol.
• Responses include the same structured provenance (sources, evidence count, contract hash) Gemini Enterprise users would see calling Asha directly.
• Authentication: Asha's own Bearer API key (per the securitySchemes.bearer declaration). Each registering org provisions its own key via POST /api/a2a/signup.

Security notes (per Google's docs)

• Gemini Enterprise's Model Armor settings in the Cloud console do not automatically protect A2A agents. If you want Model Armor on top of Asha's existing fiduciary medical contract, add it via the Model Armor REST API in your registering app.
• Asha already enforces a fiduciary medical contract (no prescribing, no definitive diagnoses, jailbreak detection, contract hash on every response) at the API boundary, independent of Gemini Enterprise.
• API keys live in your environment, never in the registered agent card. Rotate any time with POST /api/a2a/rotate-key.

Account & Operations Endpoints

In addition to the A2A protocol surface, Asha exposes operator endpoints for managing your API key and verifying corpus state:

Endpoint	Method	Purpose
POST /api/a2a/signup	None	Self-service API key provisioning
GET /api/a2a/usage	API key	Daily/monthly usage counters and limits
POST /api/a2a/upgrade	API key	Stripe checkout for tier upgrade
GET /api/a2a/portal	API key	Stripe Customer Portal (billing self-service)
POST /api/a2a/rotate-key	API key	Rotate (regenerate) your API key
GET /api/audit/corpus-state	JWT	Cryptographic corpus commitment (verifiable knowledge state)
POST /api/feng/falsify	JWT	Popperian claim falsification — see feng-a2a

Knowledge Coverage (Asha's slice)

Asha is the medical-intelligence agent. It is bound to medical, clinical, pharmacology, and research collections from the Citadel corpus:

Category	Vectors	Key Sources
Research & Academic	~90M	OpenAlex top-cited (16.5M), PubMed abstracts (5.1M), PMC full-text cited (5.2M), pkg2_abstracts (48M)
Medical & Clinical	~18M	Wikidata medical (10.9M), DailyMed drug labels (889K), StatPearls clinical (76K)
Pharmacology	~928K	DailyMed drug labels, FDA drug labels
Coding & Classification	~304K	ICD-10, structured medical codes

Total Citadel corpus across all 11 public agents: 121M+ vectors across ~591 live collections (audited 2026-05-01).

Safety

• Fiduciary medical contract enforced on every query
• Never prescribes medications or provides dosing
• Never gives definitive diagnoses
• Emergency escalation for urgent clinical scenarios
• Jailbreak detection at the API boundary
• Every response carries a verifiable contract hash

Pricing

Tier	Monthly	Daily Cap	Monthly Cap	Get Started
Free	$0	10 queries	50 queries	POST /api/a2a/signup
Developer	$49	200 queries	1,000 queries	POST /api/a2a/upgrade
Pro	$199	2,000 queries	10,000 queries	POST /api/a2a/upgrade
Enterprise	Custom	Unlimited	Unlimited	dnai.systems

Other DNAi Agents

Asha is one of 11 public agents in the DNAi fleet. All agents share the same base URL (https://api.askasha.org), auth system, and A2A protocol surface. The agent_id in metadata selects which agent handles your query.

Agent	Domain	Repo	Agent Card
Asha	Medical intelligence	asha-a2a	?agent_id=asha
Harley	Fitness coaching	harley-a2a	?agent_id=harley
Artha	Financial analysis	artha-a2a	?agent_id=artha
Sage	Nutrition & wellness	sage-a2a	?agent_id=sage
Polymath	Math & science	polymath-a2a	?agent_id=polymath
Lyra	Medical research	—	?agent_id=lyra
Leo	Legal aid	—	?agent_id=leo
Mira	Marketing & growth	—	?agent_id=mira
Ren	Customer support	—	?agent_id=ren
Arohi	Practice management	—	?agent_id=arohi
Ray	Platform architecture	—	?agent_id=ray

Companion services:

Service	Repo	Purpose
FENG (Falsification Engine)	feng-a2a	Popperian claim falsification with E-value scoring

Links

• Agent Card: https://api.askasha.org/.well-known/agent-card.json?agent_id=asha
• Fleet Card: https://api.askasha.org/.well-known/agent-card.json
• Product: https://askasha.org
• Company: https://dnai.systems
• A2A Protocol: https://a2aproject.github.io/A2A/

License

The A2A interface specification and examples in this repository are available under the MIT License. The underlying Asha engine, knowledge collections, CIU architecture, and proprietary systems are not open source.