Skip to content

Research & Engineering

Research for reliable intent-to-execution systems.

Synve's research and engineering work focuses on how AI systems can interpret unstructured intent, maintain useful context, support reviewable decisions, and move work toward reliable execution.

We are particularly interested in systems that remain understandable, bounded, privacy-conscious, and under human control.

Engineering notes

Operational evidence, methodology, and safeguards.

Evaluation

How Vortyx reduces false positives in follow-up and commitment detection

A research and engineering note on ambiguity, evidence traceability, review-first mitigation, and the evaluation evidence required before publishing quantitative claims.

Read the note

Security

How Vortyx protects security, privacy, and trust

How Vortyx protects connected workflows with data minimization, encryption, workspace-scoped access, credential separation, review-first AI, and user-controlled disconnect and deletion paths.

Read the note

Systems reliability

Scaling Vortyx without losing the thread

How Vortyx treats scalability as continuity reliability through bounded processing, semantic deduplication, durable workflow state, graceful degradation, and privacy-conscious observability.

Read the note

Intent modeling

From conversation to action: a taxonomy of work signals

A practical taxonomy for separating follow-ups, commitments, waiting items, next steps, reminders, recurring responsibilities, and drafts.

Read the note

Continuity

Waiting-on detection: identifying work that depends on someone else

How Vortyx distinguishes waiting states, ownership, timing cues, and external dependencies from ordinary unresolved conversation.

Read the note

Explainability

Evidence traceability in AI-generated follow-ups

Why AI-generated follow-ups need source, owner, timing, state, and resurfacing context before users decide whether to act.

Read the note

Human control

Why review-first AI matters for email and message workflows

Why Vortyx keeps suggestions, drafts, follow-ups, commitments, and scheduling actions reviewable before they become tracked work or external communication.

Read the note

Research areas

Applied work behind reliable action.

Intent and commitment modeling

Represent unstructured signals as actions, commitments, goals, constraints, ownership, timing, and expected next moves.

The challenge is not only identifying what was said, but determining what remains unresolved, who is expected to act, and what evidence supports that interpretation.

State and continuity

Model how unfinished work moves between action, waiting, delay, recurrence, and resolution.

Continuity requires preserving the relevant source, last meaningful activity, expected timing, current state, and reason an item may need attention now.

High-signal detection

Distinguish meaningful open loops from ordinary activity.

Our work combines model-based interpretation, confidence thresholds, deterministic checks, deduplication, and explicit review to reduce noise.

Reviewable decision support

Generate suggestions that remain visible, editable, and reversible.

The goal is to help users understand why something matters, assess the evidence, choose the next move, and retain control over external actions.

Controlled execution

Translate structured intent into bounded actions across workflows, communication systems, scheduling tools, and automation pipelines.

Execution should remain scoped, reviewable, observable, and attributable to a clear user decision.

Context, retention, and privacy

Maintain enough context to support continuity without retaining more data than necessary.

This includes user-scoped context, configurable access, minimization, deletion paths, and explicit boundaries around connected sources.

Applied product

How this work appears in Vortyx

Vortyx is the first applied product built from these research areas. It detects possible follow-ups, commitments, waiting items, recurring responsibilities, and next steps; helps users review what should be tracked; maintains the state and context of unfinished work; and resurfaces it when attention may be required.

External actions remain user-approved.

Evaluation

Reliability and evaluation

We evaluate these systems across several dimensions:

Signal quality and false-positive reduction
State and ownership accuracy
Evidence traceability
Timing and resurfacing relevance
Consistency under repeated inputs
Privacy and access boundaries
User control over suggested actions

Our objective is not maximum automation. It is dependable assistance that helps work move forward without making important decisions less visible.

Exploratory direction

Applied autonomy research

Separately from Vortyx, Synve explores how intent modeling, context maintenance, decision support, and controlled execution may extend across software and physical environments.

This work is exploratory and follows the same reliability-first principles: bounded action, observable state, explicit constraints, and human oversight.

Follow the notes.

We publish short research and engineering notes on intent modeling, continuity, extraction, evaluation, privacy, and controlled execution.