Program Analysis Product
DARPA program offices
DARPA program analysis sites: evidence you can audit
Per-program analytical sites where every quantitative claim links to a reproducible source and a stated confidence level.
See the engagementKnow where to act
when the field is too complex to read.
Nobody can read all of science. There are 225 million papers, and even an expert knows only one corner. But someone still has to decide where the next dollar, or the next year of work, should go. Syntheos draws the map. It shows who is working on what, how ideas move between groups, and where nobody is looking yet. You can check every line against its source. The same approach works for strategic assessment and in the classroom.
Running today at DARPA, the Andrew W. Marshall Foundation, and Georgetown.
Scroll
0.0BCitation Links Mapped
The citation record under the Map of Science. We follow every link between papers in the global scientific literature to find its research communities.
0+Audited Claims
Every deployed system ships with an evidence ledger. Each claim links to the source query that produced it, and the figure reruns live against the warehouse on demand. Auditors can pull any number at random. Regulators can verify the math without our help.
0+Peer-Reviewed Publications
Authored by the Syntheos team across science mapping, AI policy, computer security and privacy, and research security, including work in top-tier venues like Nature, PNAS, USENIX Security, ICLR, FAccT, and Quantitative Science Studies.
0Black Boxes
Every output traces to source evidence through W3C PROV lineage. No opaque models. No silent fallbacks.
Trusted by leaders in defense, healthcare, research, and public policy
Work
Systems we've shipped.
Every engagement ends with a working system, not a slide deck. Each of the four below solved a different shape of problem.
Decision Platform
Andrew W. Marshall Foundation
Deepfield: a modular assessment platform
An eleven-stage strategic assessment pipeline that turns a query into a defensible course of action with full evidence lineage.
See the engagement
Human-AI Teaming System
Georgetown School of Foreign Service (SEST)
Georgetown Wicked Problems Lab: a classroom where AI assists but never decides
A teaming platform in classroom use at Georgetown. The AI listens quietly and helps on request, while phase gates, peer review, and a twelve-rule delegation contract keep every real judgment in student hands.
See the engagement
Research Intelligence Tool
University of Michigan
Pathfinder: research pathway navigation for a map of science
A research pathway explorer that lets users chart and narrate routes between research communities on a global map of science.
See the engagementGovernment & DefenseBuilt for federal
Built for federal
acquisition.
Decision systems built to survive Congressional oversight and meet the standards the DoD AI office is setting.
View Defense CapabilitiesSAM.gov RegisteredNIST 800-171 In ProgressDFARS 7012 In ProgressResponsible AI by Design
Past Performance
DARPA, CSET, Marshall Foundation, Noble Reach
Mission Areas
Competitive Assessment, Technology Scouting, Decision Architectures
Acquisition Vehicles Performed
Direct Award, OTAs, BAAs
What we build
Four systems we built
from scratch.
Each of the four began with a decision someone had to make and stand behind, in a field too large to read unaided. They're original systems, built for clients who couldn't find what they needed on the market. Your engagement will take its shape from your decision. The kind of problem stays the same.
01
Program Analysis Products
Fixed-scope analytical sites where every quantitative claim ships with a reproducible query and a confidence level. Built for R&D funders who have to defend portfolio decisions on the record.
Proof: DARPA program offices
- Deployed analytical site, one per program
- Structured evidence ledger behind every claim
- Reproducible SQL queries against your warehouse
- Python refresh pipeline for standing data
02
Decision Platforms
An eleven-stage assessment pipeline you run yourself. Every stage carries W3C PROV lineage, so any final recommendation traces back through the whole chain to the source evidence. Built for standing strategic-assessment and competitive-intelligence work.
Proof: Deepfield
- Modular pipeline, eleven stages, tuned to your domain
- Knowledge graph, ACH-based reasoning, wargaming
- Adaptive compute budgets per stage
- Full evidence lineage on every recommendation
03
Research Intelligence Tools
Research consoles over specialist corpora. A researcher asks a question and gets an answer cited to the page, with figures that separate what the sources say from what the model inferred. We build a knowledge graph over the corpus when the questions need one. Built for institutions that want researchers to think with a corpus as well as read it.
Proof: Andrew W. Marshall Foundation
- Answers cited to document and page
- Hybrid vector and full-text retrieval, reranked
- Cited versus inferred marking in every figure
- A knowledge graph where the corpus calls for one
04
Human-AI Teaming Systems
Orchestrated AI assistance with human judgment in the loop. A delegation contract constrains what the AI is allowed to do, and database-level phase gates enforce it. Built for learning environments and decision-critical work where the AI must help but never decide.
Proof: Georgetown SEST
- Agent orchestration across fast, deep, and QA tiers
- Delegation contract constraining AI authority
- Database-enforced phase gates
- Voice and text surfaces, instructor tooling
Different problems, different infrastructure. Each engagement ends with a working system your team runs, not a document your team files.
Where to start
Start with
two weeks.
Send us one decision and the material behind it. We run our methods on it and walk you through the result. There's no fee.
By the end of the walkthrough you'll know whether we showed you something you couldn't see before, and we'll give you a fixed price for the full engagement.
Send us a decisionYou sendIn two weeks you get
A research area or a portfolio
A map of the research communities it touches, and where the open ground is.
Read the business caseA program name or a BAA
A live demonstration site running real analysis on the program.
Read the business caseOne strategic question
The charter, intake, and research stages of a Deepfield assessment.
Read the business caseFifty documents from your corpus
The research console answering questions on your material, cited to the page.
Read the business caseOne phase sequence you'd want to pilot
A configured teaming platform with two or three agents tuned to your domain.
Read the business caseWe take one on when the decision is consequential, the structure behind it is hard to see, and the person who owns it will be in the room for the walkthrough.
Who we work for
Where decisions
don't tolerate hand-waving.
These are the people who own the decisions we work on, in the domains the team came up through. The question changes from one to the next. The job stays the same. We build the system that shows where to act and lets the choice survive scrutiny.
Federal R&D Funders & Program Offices
Which programs to start and which to end. The portfolio then has to be defended to Congress, the press, and the next administration. We bring quantitative portfolio analysis, science mapping, growth-trajectory forecasting, and AI-assisted research-opportunity discovery to the call before it's made and to the audit that follows.
Explore capabilitiesFoundations & Strategic Philanthropy
Where a single grant will move a research community, and where it will vanish. Foundations make that call over long horizons, and the decision has to outlast the staff who made it. We've built the systems behind specialist corpora, mapped funding portfolios across decades, and modeled where a grant moves a field.
Explore capabilitiesMedical Schools & Biomedical Funders
Which lines of translational research to back, and what the clinical evidence supports. We do translational research forecasting, research-funding portfolio analysis, and clinical-evidence synthesis for medical schools, biomedical agencies, and clinical specialty boards. It's built on years of running a research-intelligence operation inside an academic medical center.
Explore capabilitiesAI & Emerging Technology Policy Teams
Which emerging technologies deserve attention now, and where research security is exposed. We do AI research-trend analysis, research-security assessment, and long-horizon emerging-technology forecasting, informed by years of work at the institutions that set this policy. The vocabulary, the methodology, and the relationships are our own.
Explore capabilitiesDefense & National Security
Where an advantage can be built or lost over a long competition. Competitive assessment demands synthesis across military, economic, and technological domains over long time horizons. Traceable reasoning. Explicit assumptions you can challenge, test, and refine.
Explore capabilitiesOutside These Domains
The domain matters less than the shape of the decision. We ask three questions. Does the decision carry real money or mission consequence? Is the structure that should drive it hidden from the people making it? Could seeing that structure change the answer? Three yeses and we should talk.
Explore capabilitiesThe team
Small group.
Deep bench.
Syntheos is a small firm by design. Every engagement is staffed by principals — people who came here from the places where high-stakes decisions are actually made.
The Office of Naval Research. The Office of the Secretary of Defense. The MacArthur Foundation. Sandia. The Naval Research Laboratory. CSET. The University of Michigan Medical School. Between us we have decades of peer-reviewed work in science mapping, research evaluation, science policy, AI policy, computer security and privacy, and research security.
You don't get handed off to a junior team. You get the people whose names are on the papers.
“
“Every decision faces two questions: was it right, and how do you know? Everything we build is for both.”
Caleb Smith
Founder & CEO
Our Approach
How an engagement runs.
01
Research
Before architecture, the investigation. We study the decision your team has to make and defend, its evidence base, the failure modes it has historically been caught by, what the literature already settles and what it doesn't. The work product belongs to you.
02
Design
The architecture follows from the research, not from a catalog. A research funder might need a bibliometric pipeline against the portfolio. An archive might need a knowledge graph over the corpus. A strategic-assessment team might need an eleven-stage workflow with provenance lineage end to end. We design the system that fits the decision, not the other way around.
03
Implement
We build it in your environment, on your data, against the security posture your auditors already recognize. Every claim the system produces traces back through W3C PROV lineage to the source evidence. The methods stay yours when we're done.
04
Hand off
Your team runs the system without us. We document the pipelines, train your operators, and stay close enough by phone to fix anything broken in the first month. The system is yours, the data is yours, and the decisions made on top of it are yours. Sometimes that's where it ends. More often the first system shows a client a question they couldn't have asked before, and they bring us back to go after it.
FAQ
Questions From Leaders Who've Been Here Before
These are the questions that come up when we talk to leaders who've seen enough vendor pitches to be skeptical. They're worth asking of anyone in this space, including us.
AI projects often fail because they're framed as technical projects when they're actually organizational change projects. Someone buys a platform, a team builds models, the models perform well on test data, and then nothing changes. The decisions get made the same way they always did, by the same people, using the same inputs. The AI sits on the side, technically functional and organizationally irrelevant.
The usual culprit is starting from the wrong end. "We have all this data, what can we do with it?" is a recipe for building impressive demos that don't matter. So is "leadership wants an AI initiative." These projects optimize for what's measurable by technologists, like model accuracy, processing speed, and data coverage. Decision-makers care about whether they can trust an output, explain it to their boss, and defend it when something goes wrong. F1 scores are not on that list.
We start from the other end. A specific decision that matters, made by specific people, who answer to specific stakeholders. What evidence do they need? What does defensible look like in their world? Where does their current process break down? The technical architecture follows from those answers, not the other way around.
This is less exciting than a general-purpose AI platform. It's slower to show results in a demo. But it's the difference between a system that gets used and a system that gets abandoned.
Initiate Contact
Bring us a decision you have to make and defend.
Tell us about the decision you're trying to improve. We'll schedule a briefing with our principals to understand your environment and see whether the fit is right.
Schedule a briefing