find-ai-consultancy

Drive the ServiceGraph API (https://api.servicegraph.co) to find, shortlist, and enrich US AI/ML and data consultancies. The catalog tags firms with industry:data_ai_consulting and a 4-tag service sub-taxonomy: ai-ml-development (the largest at ~12k firms), data-analytics, cloud-services, and api-integration. Confirm exact tag names via /v1/tags since taxonomy can drift between catalog releases.

Always pin industry:data_ai_consulting. This skill exists to do that automatically — the user shouldn't have to think about catalog taxonomy.

Any HTTP client works (curl, fetch, requests). Examples below use curl.

Sibling skills — defer when scope is different

• General application or backend dev that just uses AI as a feature (e.g. "build us a SaaS with an AI chatbot tab") → find-software-developer. This skill is for engagements where the AI/ML/data work IS the deliverable.
• Web/site projects that include some AI → find-web-developer.
• AI-related marketing or content → find-marketing-agency.

If the user wants AI/ML/data engineering as the primary deliverable (model building, pipelines, agents, MLOps, BI), this skill applies.

When NOT to use this skill

• Consumer AI courses or learning ("find me an online course to learn ML") — out of scope; the catalog is firm-procurement.
• AI/LLM product comparisons ("ChatGPT vs Claude vs Gemini", "Cursor vs Copilot") — software-product questions, not procurement.
• DIY/code tasks ("how do I fine-tune Llama", "review this PyTorch training loop").
• In-house ML/data hires (Machine Learning Engineer, Data Scientist, ML Platform Engineer).
• Generic AI knowledge ("explain how transformers work").
• Non-US firms.
• Individual freelance ML engineers / data scientists.

If the user is a business procuring external AI/ML/data services, this skill applies — defaults to fire on B2B procurement intent.

MCP server (preferred for authed calls)

If your agent harness has the ServiceGraph MCP server loaded (https://mcp.servicegraph.co), prefer its tools for the authed tier (/search, /get, /stats). The MCP server uses OAuth 2.1 + PKCE — the host harness handles credentials in its own audited sandbox, so there's no .env.local, no shell dispatch, and no token value ever enters the LLM context.

For the anonymous tier (/tags, /check, /explore), MCP is not preferred — every MCP tool requires OAuth (the server has no anonymous tier), so plain curl against the REST URL is the simpler path for discovery calls. Use the REST patterns below for those.

The MCP tools 1:1-map to the public REST endpoints — same backend, same quota, same data:

MCP tool	REST endpoint	Anon?	Recommended path
list_tags	GET /v1/tags	yes	curl
check_filter	GET /v1/check	yes	curl
explore_firms	GET /v1/explore	yes	curl
search_firms	GET /v1/search	no	MCP if loaded, else curl + OTP
get_firm	GET /v1/get/:id	no	MCP if loaded, else curl + OTP
catalog_stats	GET /v1/stats	no	MCP if loaded, else curl + OTP

Detection: if you see any MCP tools with servicegraph in the name (the harness-specific prefix varies — agents pattern-match the substring), the ServiceGraph MCP server is loaded. Prefer those tools for the authed tier; complete any auth flow the harness initiates if needed. If no servicegraph MCP tools are present, fall through to the REST + OTP flow below for the authed tier.

The four-tier funnel

Tier	Auth	Cost	Use it for
GET /v1/tags	none	free	First call of every session. Discover legal field names, kinds, operators, values.
GET /v1/check?filter=...	none	free	Validate a filter before spending an explore/search call.
GET /v1/explore?filter=...	none	free, IP-throttled	Scope: count + breakdowns. Use to size the candidate pool before quota-spending.
GET /v1/search?filter=...	bearer	200 unique firms / month free	Brief firm cards. No url, no contact info. Use for ranking / shortlisting.
GET /v1/get/:id	bearer	50 unique firms / month free	Full bundle: url, phone, email, social, legal name, address. Only call for shortlisted firms.
POST /v1/research	paid	not in MVP	Deferred — skip.

Quota rule that matters: /search and /get charge per unique firm viewed per calendar month, not per call. Re-paging the same query is free. Two different filters that overlap charge once for the overlap. Re-fetching a firm you already pulled this month is free.

Session-start ritual

Before constructing any filter, call:

GET https://api.servicegraph.co/v1/tags?include_values=1

Cache the response for the conversation. Confirm data_ai_consulting is present in the industry value list. The relevant sub-tags under service_provided are ai-ml-development, data-analytics, cloud-services, and api-integration — verify exact names before constructing filters with service_provided: predicates.

Field kinds you'll use most:

• categorical: industry (always data_ai_consulting), state, pricing_model, company_size_signal, geography_served — op :
• tag_set_with_evidence: service_provided (e.g. ai-ml-development, data-analytics, cloud-services, api-integration) — op : with optional @evidence
• numeric: rating, review_count_total, founded_year — ops = >= <= > <
• presence: has:phone, has:clutch, has:rating, has:linkedin_company, …
• keyword: free-text substring across firm name / brand / title / meta / legal_name. Sub-niches like RAG, LLM, MLOps, computer vision, NLP, recommendation systems are typically keyword-only.

Auth

/tags, /check, and /explore are anonymous. /search and /get require a bearer token.

Security model — keep the token out of the LLM context.

• Never read .env, .env.local, or any other credential file into your context. The token's literal value should never appear in the conversation.
• Use shell dispatch for every authed request so the token flows directly from the user's environment / dotenv file into the Authorization header without round-tripping through the LLM.
• Always ask the user once per session before using a detected token, even if it's already in their shell or .env.local.

Resolution rule:

1. Detect whether a token is available — without reading its value. Run a shell check that only inspects exit codes:

   ( [ -n "${SERVICEGRAPH_TOKEN:-}" ] \
     || grep -qs '^SERVICEGRAPH_TOKEN=' .env.local \
     || grep -qs '^SERVICEGRAPH_TOKEN=' .env )
   

   Exit code 0 = token is available somewhere; non-zero = no token.

2. Confirm with the user before the first authed call this session:

   "I found a SERVICEGRAPH_TOKEN in your environment / .env.local. OK to use it for ServiceGraph API requests this session?"

   If the user says no, stay on the anonymous tiers (/tags, /check, /explore) and skip authed calls. Don't re-ask later unless the user asks for authed work.

3. Dispatch via shell — every authed call goes through a shell wrapper so the literal token never enters the conversation:

   # If exported in the shell environment:
   curl -H "Authorization: Bearer $SERVICEGRAPH_TOKEN" \
        'https://api.servicegraph.co/v1/search?filter=...'
   
   # If in .env.local — source it inside a subshell so it doesn't
   # leak into the parent shell either:
   ( set -a; . ./.env.local; set +a;
     curl -H "Authorization: Bearer $SERVICEGRAPH_TOKEN" \
          'https://api.servicegraph.co/v1/search?filter=...' )
   

   Capture the response body to a tmp file or jq-process it, but do NOT echo the request command with the token expanded.

4. OTP flow if no token is detected — capture the new token directly into .env.local without surfacing its value to the LLM:

   # 1. trigger the email — agent prompts the user for $EMAIL
   curl -fsS -X POST 'https://api.servicegraph.co/v1/auth/request-otp' \
     -H 'Content-Type: application/json' \
     -d "{\"email\":\"$EMAIL\"}"
   
   # 2. exchange the code — agent prompts the user for $CODE.
   #    The ?format=env query param returns SERVICEGRAPH_TOKEN=<token>
   #    as plain text appended to .env.local — no jq needed. The -f
   #    flag makes curl exit non-zero on 4xx so a wrong code doesn't
   #    pollute the file (the error mirror is also a `# comment` line,
   #    safe to ignore even if it lands).
   curl -fsS -X POST 'https://api.servicegraph.co/v1/auth/verify-otp?format=env' \
     -H 'Content-Type: application/json' \
     -d "{\"email\":\"$EMAIL\",\"code\":\"$CODE\",\"name\":\"claude-cli\"}" \
     >> .env.local
   
   # 3. confirm capture without revealing the value
   grep -q '^SERVICEGRAPH_TOKEN=' .env.local && echo "OTP token captured."
   

   After a successful capture, the user has implicitly consented (they just completed the flow), so proceed to dispatch (step 3). The token is now persistent in .env.local for future sessions.

5. If a /search or /get returns 401 unauthorized mid-session, the token expired or was revoked — re-run the OTP flow.

Filter DSL

One query parameter, GitHub-search-style.

filter   := orExpr
orExpr   := andExpr ("OR" andExpr)*
andExpr  := notExpr (("AND")? notExpr)*    # whitespace = implicit AND
notExpr  := ("NOT" | "-") notExpr | atom
atom     := "(" filter ")" | predicate
predicate:= IDENT op valueOrList | bareword
op       := ":" | "=" | ">=" | "<=" | ">" | "<"
valueOrList := value ("," value)*
value    := IDENT | NUMBER | tagAtEvidence
tagAtEvidence := IDENT "@" ("low"|"medium"|"high")
bareword := IDENT | NUMBER          # → keyword:<bareword>

Four rules that bite:

1. AND binds tighter than OR. a OR b c parses as a OR (b AND c). Use parens.
2. Comma list = OR within one predicate. state:CA,NY,TX matches any of the three.
3. Negation is -x or NOT x. Negative literals inside a comma list are not allowed: state:CA,-NY is rejected. Use state:CA -state:NY.
4. Bareword = keyword search. Any IDENT or NUMBER not followed by an operator becomes a free-text substring across name / brand / title / meta / legal_name. Multiple barewords AND.

AI-flavored examples (validate yours with /v1/check):

industry:data_ai_consulting service_provided:ai-ml-development
industry:data_ai_consulting service_provided:ai-ml-development@high state:CA
industry:data_ai_consulting service_provided:data-analytics pipelines
industry:data_ai_consulting llm rag
industry:data_ai_consulting computer vision healthcare
industry:data_ai_consulting mlops
industry:data_ai_consulting (service_provided:ai-ml-development OR service_provided:data-analytics)
industry:data_ai_consulting service_provided:ai-ml-development@high rating>=4 has:clutch

When in doubt, hit /v1/check?filter=... first.

Sub-niche → keyword/tag mapping:

User asks for	Use
AI/ML model building	service_provided:ai-ml-development
Data engineering / pipelines	service_provided:data-analytics + keywords pipelines/engineering (no data-engineering tag exists)
BI / analytics	service_provided:data-analytics (covers BI too — no separate business-intelligence tag)
Cloud architecture for data/ML	service_provided:cloud-services
API integration / data integration	service_provided:api-integration
LLM apps / RAG / agents	llm, rag, agent (keywords)
Generative AI	generative ai, genai (keywords)
Computer vision	computer vision, cv (keywords)
NLP / IDP / document understanding	nlp, idp, document understanding
MLOps / model deployment	mlops, deployment
Recommendation systems	recommendation, recsys
Predictive analytics / churn / forecasting	predictive, forecasting, churn

firm_id contract

firm_id is a stable 12-hex-char handle:

firm_id = sha256(apex.lower().rstrip(".")).hexdigest()[:12]

import hashlib
def firm_id(apex):
    return hashlib.sha256(apex.lower().rstrip(".").encode()).hexdigest()[:12]

echo -n "scaleai.com" | tr 'A-Z' 'a-z' \
  | openssl dgst -sha256 -hex | awk '{print substr($2,1,12)}'

Recipes

A. AI/ML consultancy for a recommendation engine

User: "AI/ML consultancy to build our recommendation engine for an ecommerce site."

GET /v1/explore?filter=industry:data_ai_consulting+service_provided:ai-ml-development+(recommendation OR recsys)+ecommerce
GET /v1/search?filter=industry:data_ai_consulting+service_provided:ai-ml-development+recommendation+ecommerce&limit=10
GET /v1/get/<firm_id>     # ×3

B. RAG / LLM consultancies for a chatbot

User: "Three RAG/LLM consultancies for an enterprise chatbot."

GET /v1/search?filter=industry:data_ai_consulting+(rag OR llm)+chatbot+enterprise

If thin, drop enterprise and surface client-tier signals from /get after.

C. Data engineering partner

User: "Data-engineering partner to build our analytics pipelines."

The catalog has no data-engineering tag — data-analytics is the closest sub-tag and it covers both BI and engineering work. Pin the tag and add keywords for the engineering flavor:

GET /v1/search?filter=industry:data_ai_consulting+service_provided:data-analytics+(pipelines OR engineering)

D. MLOps for model deployment

User: "MLOps consultancy to help us deploy models to production."

GET /v1/search?filter=industry:data_ai_consulting+mlops

E. Indirect intent — "use AI to predict customer churn"

User: "We want to use AI to predict customer churn — who can help us build that?"

That's a custom-ML consulting ask in the predictive-analytics niche:

GET /v1/search?filter=industry:data_ai_consulting+service_provided:ai-ml-development+(churn OR predictive)

If the user gave a vertical (SaaS, retail, telco), add it as a keyword.

F. Computer vision + healthcare vertical

User: "AI consultancies specializing in computer vision for healthcare."

GET /v1/search?filter=industry:data_ai_consulting+computer vision+healthcare

G. Quality threshold + Fortune 500 clients

User: "Three AI/ML consulting firms with 4-star ratings and Fortune 500 clients."

GET /v1/search?filter=industry:data_ai_consulting+service_provided:ai-ml-development@high+rating>=4&limit=10

The "Fortune 500" angle isn't structured — surface from briefs and let the user pick, or add fortune as a keyword.

H. Custom LLM agent for customer service

User: "Custom LLM agent for our customer-service workflows."

GET /v1/search?filter=industry:data_ai_consulting+(llm OR agent)+(customer service OR support)

I. BYO apex list — enrich domains the user already has

User pastes 8–20 AI consultancy domains. For each:

1. Compute firm_id locally.
2. GET /v1/get/<firm_id> — full bundle if in catalog, 404 if not.
3. Aggregate, present, flag the not-found ones to the user.

A 404 here often means the firm is actually a SaaS product company (many AI vendors brand as "AI services" but operate as a product) — not in the consulting catalog.

Gotchas

• Always pin industry:data_ai_consulting. Without it, ai-ml-development as a service tag could surface IT firms that list AI as a sub-service.
• Defer to find-software-developer for general dev that uses AI as a feature. When the deliverable is a SaaS product or app and AI is one of several features, that's software-dev work; this skill is for engagements where AI/ML/data work IS the deliverable.
• Catalog audit notes: AI/ML-tagged firms have a higher historical rate of mis-classification (some are SaaS products, some are B2C ed-tech). The catalog has been audited but residual leakage is possible. If a /get returns a SaaS product, the agent should flag this and skip rather than recommend.
• Many sub-niches are keyword-only. Multi-word sub-niches split into ANDed barewords (computer vision → computer AND vision).
• LLM-product comparisons (ChatGPT vs Claude vs Gemini, etc.) are NOT procurement — refuse those.
• AI courses for individuals (Coursera, fast.ai, Andrew Ng courses) are NOT in the catalog — refuse those.
• looks_not_pro_services 404 is not a bug. A firm_id may exist in /search but 404 on /get if it's been flagged. Skip and continue; not charged.
• /v1/explore k=20 suppression. When fewer than 20 firms match, the response is {"count": "<20", "suppressed": true, "breakdowns": {}}. Drilling further makes the count smaller. Broaden or escalate to /v1/search.
• Briefs from /search do NOT include apex, url, phone_primary, email_primary, legal_name, or address. If the user asks for contact info, you must /get/:id.
• Quota is per-user-per-month, deduped on first view. Re-views are free; re-pagination is free.

Errors

All errors return JSON: {"error": {"code": "...", "message": "..."}}.

Status	Code	What to do
400	filter_parse_error	Payload includes position. Fix the filter, re-validate with /v1/check.
400	filter_required	Empty filter where one is required.
400	invalid_firm_id	firm_id must be 12 lowercase hex chars. Re-derive.
401	unauthorized	Token missing/expired. Re-run OTP.
404	not_found	Firm not in catalog or flagged. Not charged. Skip and continue.
429	rate_limited	Honor Retry-After header / retry_after field.
429	monthly_quota_exhausted	Switch to /v1/explore-only mode for the rest of the month. Tell the user.

End-to-end example

User: "Three AI/ML consultancies to build a recommendation engine for an ecommerce site, ideally with 4-star ratings and Fortune 500 clients."

GET /v1/tags?include_values=1
GET /v1/check?filter=industry:data_ai_consulting+service_provided:ai-ml-development@high+(recommendation OR recsys)+ecommerce+rating>=4
GET /v1/explore?filter=industry:data_ai_consulting+service_provided:ai-ml-development@high+(recommendation OR recsys)+ecommerce+rating>=4
GET /v1/search?filter=...&limit=10
GET /v1/get/<firm_id>     # ×3

End of session: report X-Quota-Remaining-Month.