Research
We built the model. This is how.
Syntic Research is the team that built Amara from the floor up — and keeps building it. Pretraining. Alignment. Evaluation. Multi-agent coordination. Voice. Long-horizon autonomy. Safety under real adversarial pressure. Every line of research ships directly into the platform our customers run in production. Not into a deck. Not into a press release. Into the product.
Why we built the lab instead of licensing a model
Most AI companies don't have a research lab
Most companies calling themselves AI companies don't have a research lab. They have an integration layer. They license a model from OpenAI or Anthropic, wrap it in a product, raise venture capital, and describe themselves as frontier AI. The actual research — the pretraining, the alignment, the architecture decisions — happened at someone else's company, on someone else's compute, and belongs to someone else entirely.
We made a different decision. Building Amara from scratch required real research investment: architecture, tokeniser design, training data curation, alignment methodology, evaluation infrastructure, safety layers, voice synthesis, multi-agent coordination. We absorbed that cost because the alternative was to build a business on a foundation we couldn't see, couldn't control, and didn't own. That isn't a business. It's a dependency.
The lab isn't a marketing function. It isn't a recruiting signal. It is literally how Amara exists and how the Syntic platform improves. Cut the research team and Amara stops getting better. That's not true at most AI companies — at Syntic, it is.
We publish what we learn. We ship what we publish. The pipeline runs in weeks, not years.
The research threads
Six areas. All active. All shipping into production.
The science underneath Amara and the AI workforce platform, from pretraining to production safety.
Foundation model research, Amara
The training of Amara itself — from architecture decisions and tokeniser design through data curation, post-training alignment, inference optimisation, and scaling laws calibrated to our own compute envelope. Amara is live in production across both Syntic products today. The next-generation pretraining run is underway. This is not exploratory research. It is the work that determines what ships.
Alignment and safety
How a model behaves the way its designers and users intend — under normal use, under adversarial pressure, and across multi-day autonomous deployments. Our work spans RLHF and successor techniques, constitutional methods, and red-team-driven hardening. The safety properties we research are specific to AI workforce deployments: not just chat safety, but the safety of an AI Employee that books meetings, sends payments, and modifies production systems.
Long-horizon agent autonomy
The hardest open problem in production AI — and the one that determines whether AI Employees are real. How does an AI Employee maintain coherent intent across a multi-day task? How does it handle drift, surface uncertainty when the situation has changed, recover from a mistake, and stay inside its authority? Our autonomy threshold framework defines four measurable levels — from supervised assistant to autonomous executor — with the runtime guarantees we ship at each. This isn't a whitepaper exercise. These thresholds gate every production deployment in Syntic Workforce.
Multi-agent coordination
How groups of AI Agents work together under a named Employee. Supervisor-worker patterns, peer-to-peer negotiation, market-style dispatching when work is parallelizable, and the Council pattern for deliberative second opinions on high-stakes decisions. We've published patterns from production Workforces handling concurrent dispatches at scale.
Evaluation
How you measure whether a non-deterministic AI workforce is actually working. Our evaluation harnesses gate every deploy of the Syntic platform. We've developed methodology around golden datasets versus synthetic ones, regression tiers, behavior diffs across model versions, and continuous evaluation during production. The evaluation work is the bridge between research and ship.
Voice and multimodal
The infrastructure powering the AI Call Center. Real-time voice synthesis with sub-second latency, custom voice cloning, barge-in and natural conversational rhythm, and voice in 16 production languages. The multimodal direction, voice plus vision plus text, is integrated into Amara's next-generation training.
Recent and selected work
Open publications from the Syntic research team
A selection of the work we've put into the open, from autonomy and evaluation to voice and adversarial safety.
Autonomy thresholds in long-horizon AI Employee tasks
A framework for measuring when an AI Employee transitions from supervised assistant to autonomous executor. Empirical results across coding, research, customer service, and operational workloads. Defines four autonomy levels (L0 to L3) and the runtime safety guarantees we ship at each level.
Read paperPretraining Amara, design decisions for an independent frontier model
The architectural and training decisions behind Amara. Why we made the choices we made on tokenizer, context length, attention mechanism, and training data. What an independent lab learns when it can't just copy a competitor's approach.
Read paperEval harnesses for non-deterministic Workforce systems
How we build evaluation suites that gate every deploy of the Syntic Workforce. Why golden datasets outperform synthetic ones for production AI systems. Regression tier design. The methodology behind shipping non-deterministic systems with confidence.
Read paperCoordinating teams of Syntic AI Agents
Patterns for supervisor-worker, peer-to-peer, and market-style coordination among AI Agents. Results from production Workforces handling concurrent dispatches. When coordination breaks down and how to detect it before it costs the customer.
Read paperSafety properties of an AI Workforce under adversarial load
Red-team findings on prompt injection, lateral movement between AI Employees, sandbox escape attempts, credential exfiltration, and channel takeover. The runtime guarantees Syntic ships against each threat class. Honest accounting of what we mitigated and what we're still working on.
Read paperVoice synthesis at production latency
The infrastructure powering real-time AI Call Center voice. How we achieve sub-second response latency with barge-in across 16 languages. Custom voice cloning quality, ethical guardrails, and disclosure handling per jurisdiction.
Read paperThe Council pattern for high-stakes AI decision making
Multi-Employee deliberation for decisions where a single AI judgment carries too much risk. The pattern, the failure modes, the production results, and where it works versus where it doesn't.
Read paperCompliance-aware language enforcement in production AI agents
How we encode FDCPA, HIPAA, TCPA, and equivalent regulatory frameworks into Employee runtime behavior. Counsel-reviewed phrasing libraries. Real-time enforcement at the token level. Audit trail design for regulated workflows.
Read paperWhy we publish
The frontier is closing
The AI labs that defined this field used to publish. Attention Is All You Need. Scaling Laws. Constitutional AI. The papers that built the modern ML stack came out of labs that believed shared knowledge accelerated everyone. That tradition is eroding. Frontier work now happens behind closed doors. Researchers join labs that won't let them publish. The open ecosystem that created modern AI is being privatised by the companies that benefited most from it.
Syntic publishes because the open tradition built this field and deserves to survive. Because a researcher should be able to read, cite, and build on the work that came before them. Because a customer running production AI on our platform should be able to read the science underneath what they're paying for. And because we're independent enough — no investor relations team, no competitive moat to protect through obscurity — to make that call without a committee approving it.
We don't publish everything. Some work stays internal because it would compromise customer security or expose active attack surfaces we haven't fully closed. But the default is open, and we'll defend that default even when it's inconvenient.
How research becomes product at Syntic
The pipeline from paper to production
At most research labs, the pipeline from paper to product is measured in years. A paper ships. It influences thinking. Months later, an engineer reads it. Eventually something changes in the product. The research and the engineering are in different buildings, on different schedules, measured by different metrics.
At Syntic, research and engineering are the same team working on the same model, shipping into the same platform, measured by whether the product got better. The autonomy threshold framework from last quarter's paper is the runtime authority model running in Cowork today. The multi-agent coordination patterns in our published work are the exact patterns the AI Call Center uses to dispatch concurrent calls. The evaluation harness methodology powers every Syntic platform deploy.
The cycle from research to production runs in weeks. That isn't a feature we engineered — it's what happens when there's no wall between the lab and the product.
It also means our research has something most academic work doesn't: production grounding. We know what actually works at scale, with real customers, under real adversarial pressure, because we operate it. Benchmark performance that doesn't translate to deployments doesn't survive contact with our evaluation harnesses.
The team
Small lab inside a small company
Syntic Research is part of the 26-person Syntic team: ML researchers, alignment engineers, evaluation specialists, voice and multimodal researchers, and safety engineers. Most have shipped production AI systems at major labs or academic institutions before Syntic.
Distributed worldwide. No headquarters. We hire researchers wherever they are.
The lab publishes under its own bylines. Individual researchers are credited on the work they do. Our papers are real research papers, not marketing-driven white papers; they go through internal peer review and external academic review where appropriate.
Working with the research lab
Collaborate, test, commission, or join
Four ways researchers, red teams, enterprises, and prospective hires work with the lab.
For academic researchers
We collaborate with academic labs on alignment, safety, evaluation, and multi-agent coordination research. Compute partnerships are available for joint research projects, and researcher visits and sabbaticals at Syntic are possible for senior academics. Contact research@syntic.ai with proposals.
For safety researchers and red teams
We engage independent safety researchers and red teams on adversarial testing of Amara, the Syntic Workforce platform, and customer-facing deployments. Compensated engagements are available for significant findings. Responsible disclosure program at security.syntic.ai.
For enterprise customers
Enterprise customers commissioning custom Amara variants get direct engagement with the research lab on training data design, fine-tuning approaches, evaluation methodology, and safety properties specific to their deployment. Most enterprise custom-model engagements include research team participation.
For researchers wanting to join
We hire ML researchers, alignment engineers, evaluation specialists, voice and multimodal researchers, and safety engineers. Worldwide-distributed, remote-first, no relocation required. Competitive compensation, full publication rights on non-confidential work, and real ownership of the work you ship.
Compute and infrastructure
The research is real. The compute is real.
Syntic Research operates on our own dedicated training and inference infrastructure. Amara's pretraining runs on GPU clusters we operate ourselves; production inference runs on the same infrastructure that serves customer workloads. We don't rent compute from competitors. Independence at the model layer requires independence at the infrastructure layer.
For independent academic researchers, we provide compute grants on a case-by-case basis for alignment, safety, and evaluation research that benefits the broader field. Contact research@syntic.ai with proposals.
What we're working on next
The active research agenda
Selected highlights from the threads we're pushing on now.
Next-generation Amara
The next pretraining run is underway. Larger context up to multi-million token, native multilingual support across 40+ languages, native voice and vision capabilities, and improved reasoning across long-horizon tasks. Public release on the 2026 roadmap.
Autonomous AI workforces at scale
What happens when an AI Employee operates with minimal human supervision over weeks and months? How do drift, intent preservation, and accumulating decisions interact? How do you measure long-horizon outcome quality? An active research thread.
Custom model training for regulated industries
The science of training enterprise-specific Amara variants on proprietary data while preserving safety properties, compliance behavior, and quality. The research bridges fine-tuning, RLHF, and constitutional methods.
Multi-Employee deliberation
The Council pattern works for some classes of decisions. We're researching when it doesn't, when adding more Employees to a deliberation makes things worse, and what alternative coordination patterns work for genuinely hard collective judgments.
Voice and embodiment
The next frontier for AI Employees: physical-world presence through phone, video, and eventually robotic interfaces. Voice is the first step, multimodal voice and vision is the second, and embodied AI workforces are the long horizon.
Safety under coordination
Most AI safety research focuses on a single AI system. The Syntic platform routinely deploys teams of AI Employees and Agents working together. We're researching the safety properties of coordinated AI systems: how risks compound, how to evaluate them, and how to design coordination patterns that fail safe.
Browse our publications
Read our papers by topic
The full body of Syntic research, organized by thread.
Get in touch
How to reach the research lab
Collaboration, red teaming, academic partnerships, careers, and custom-model engagements.
For research collaboration
research@syntic.ai
For safety researchers and red teams
security@syntic.ai
For academic partnerships and compute grants
research@syntic.ai
For press inquiries about research
press@syntic.ai
The future of AI will be built in labs that own what they build, publish what they learn, and ship what they publish
Syntic is one of them.
Independent, open by default, and accountable to the customers running production AI workforces on our work. Not to a cap table. Not to a benchmark. To what actually ships.