A local-first AI agent that does repetitive patent-research chores for people who don't have a patent database or a patent lawyer.
Built for the 2026 Agents for Humans hackathon with the Strands Agents SDK · Track: Professional Agents.
📹 Watch the demo on YouTube → — every figure in it is real agent output, not a mock-up. (also in-repo)
Patent research is a repetitive chore locked behind expensive databases, cryptic query syntax, and API-key hoops. So the people who could benefit most — indie developers deciding what to build, founders sizing a space, students, journalists, small-town analysts — usually just… don't. "Who's actually filing patents in this area?" "Is this trend real or hype?" "Do we have freedom to operate?" Answering these means an afternoon of manual searching, over and over.
Patent Scout turns those repetitive lookups into a single plain-language request and answers with real numbers from public patent data — running on a local model so there's no key, no cloud account, and nothing sensitive leaves the machine.
you> Who files the most agentic AI patents? Top 5, and one surprising takeaway.
Patent Scout: Top 5 filers of agentic-AI patents (sampled):
1. Salesforce, Inc. — 71
2. Stats LLC — 37
3. Eric L. Pankey — 33
4. Citibank, N.A. — 25
5. Aieconomy LLC — 16
Surprising takeaway: it's enterprises like Salesforce and Citibank — not just AI
labs — and a small firm, Stats LLC, sits at #2. (Counts are sampled shares, not
exact totals.)
(Real output from the local agent. How many patents mention large language model?
returns ~125,048 the same way — every figure comes from a tool call, never the model.)
A Strands agent with eight tools over a real public-patent engine. The agent decides which tools to call, chains them for multi-step questions ("compare these three companies, then tell me who's rising"), and writes a short answer a non-expert can act on.
- Local & private by default — talks to a local model through Ollama (
llama3.1:8b). No API key, no cloud, reproducible by any judge. - Grounded — every number comes from a tool call against live public data; the system prompt forbids inventing counts.
- Honest — it flags that leaderboard counts are sampled and that recent years are undercounted (patents publish ~18 months after filing).
flowchart LR
U[User<br/>plain-language task] --> A
subgraph Strands Agents SDK
A[Patent Scout agent<br/>system prompt + tool loop]
M[(Local model<br/>Ollama · llama3.1:8b)]
A <--> M
end
A -->|tool calls| T
subgraph Tools patent_scout/tools.py
T[patent_count · top_filers · filing_trend<br/>compare_companies · company_portfolio<br/>recent_filings · list_domains · domain_landscape_report]
end
T --> E[patent-intel engine<br/>vendored, MIT<br/>24h cache · 2s pacing · parser]
E -->|HTTPS| G[(Google Patents<br/>public search)]
E --> T --> A --> R[Short, cited answer]
Swap the model provider with two env vars — or point Strands at AWS Bedrock / Bedrock AgentCore for a cloud deployment (see below).
pip install -r requirements.txt # strands-agents, ollama
ollama pull llama3.1:8b # any tool-capable local model works
python -m patent_scout.agent "Compare OpenAI, Anthropic, and NVIDIA on large language model patents."Run it interactively:
python -m patent_scout.agentTry:
How many US "retrieval augmented generation" patents are there, and show me 5 recent ones?Give me a quick landscape of the rag domain.(runlist domainsto see slugs)Is "humanoid robotics" patenting rising, and who leads?
Config (optional):
| Env var | Default | Meaning |
|---|---|---|
PATENT_SCOUT_PROVIDER |
ollama |
ollama or bedrock |
PATENT_SCOUT_MODEL |
llama3.1:8b |
model id (per provider) |
PATENT_SCOUT_HOST |
http://localhost:11434 |
Ollama host |
PATENT_SCOUT_REGION |
us-east-1 |
Bedrock region |
Strands is provider-agnostic, so the switch is one env var — the tools and
prompt don't change, only the model provider (build_model in
patent_scout/agent.py):
# uses your standard AWS credentials (env / profile / role)
export PATENT_SCOUT_PROVIDER=bedrock
export PATENT_SCOUT_MODEL=us.anthropic.claude-3-5-sonnet-20241022-v2:0 # or any model you can invoke
python -m patent_scout.agent "How many patents mention large language model?"This is the drop-in path to a hosted / Bedrock AgentCore deployment.
pip install pytest && python -m pytest tests/ -qThe suite covers the trust-critical logic — query normalization, the tool-call
loop guard, error wrapping, and the provider switch — with no network or model
required. CI runs it on Python 3.10–3.12 (see the tests badge / workflow).
| Tool | What it answers |
|---|---|
patent_count |
How many patents match a topic? |
top_filers |
Who files the most (leaderboard)? |
filing_trend |
Is a topic rising or cooling, year by year? |
compare_companies |
Who has more patents in X — A, B, or C? |
company_portfolio |
A company's total + most recent filings |
recent_filings |
The newest patents on a topic |
list_domains |
Built-in domain packs (agentic-ai, rag, ev-battery, …) |
domain_landscape_report |
Full landscape (volumes, filers, trend, fresh) for a domain |
For indie devs, small orgs, students, journalists, and analysts it replaces an afternoon of manual searching with one question — for free, locally. It is bibliographic analysis, not legal advice: leaderboard counts are sampled (relative shares, not exact), recent years undercount, and it does not replace a professional freedom-to-operate or novelty opinion.
This repository — the Strands agent, its tools, docs, and diagram — is new
work built during the submission period. The underlying patent engine in
patent_scout/vendor/patent_search.py
is vendored, unmodified, from my own prior open-source project
patent-intel (MIT). See
NOTICE.
MIT © 2026 Byungwon So
