Context is fragmented
CVs, job descriptions, proposals and course details live in different places.
AI Communication Agent
Your AI agent for important emails.
Give ContextMail your goal and materials. It understands the task, plans the workflow, drafts the email and verifies it before you send.
Built for high-context communication: job applications, research outreach and school affairs.
Goal receivedPlanning
User goal
“Use my materials to write a personalized PhD enquiry.”
CV.pdfProposal.pdf
01Goal
02Plan
03Research
04Write
05Review
06Approve
✓
Ready for approvalClaims reviewed · 2 sources attached
The problem
Important emails require more than writing.
The hardest part is rarely the sentence. It is understanding the goal, assembling the right context and deciding what can safely be said.
Research takes time
Important messages often require recipient or organization context before writing starts.
Generic AI lacks context
One prompt tends to produce generic language rather than a goal-specific strategy.
AI can hallucinate
Confident prose can include claims that were never supported by the source material.
Users need control
Important external communication should never leave without deliberate approval.
Interactive Product Demo
Describe the goal. Let the agent decide.
Start with an example or edit the request. The Planner identifies the intent and selects the workflow after you run it.
Try an exampleExamples fill the input — they do not manually set the final intent.
01
Your requestPre-filled scenario
DOC
CV.pdfDemo material · ready
✓DOC
Research_Proposal.pdfDemo material · ready
✓This static demo uses pre-authored mock data. No backend, external search or email provider is called.
02
Agent activityWaiting to start
Ready to plan
The planner will select only the agents this scenario needs.
Human in the loop
AI works. You stay in control.
ContextMail can understand, plan, research, draft and review. But important external actions remain a human decision.
No important external action without user approval.AI workPlan · Prepare · Draft · Review
Decision pointHuman Approval
Connected systemExternal Action
Product workflow
How ContextMail works.
One goal enters. A task-specific, evidence-aware workflow comes out.
Understand
Identify the user’s goal, recipient and communication scenario.
Plan
Decide what context, tools and specialist agents the task actually needs.
Prepare Context
Turn materials and relevant information into traceable evidence.
Generate & Review
Draft the message, then independently check quality and claims.
Approve
Show the complete action preview and wait for the user’s decision.
System design
Planner-led, not pipeline-bound.
The workflow adapts to the task, shares traceable state and loops back when review finds a problem.
InputUser Goal + Materials
↓OrchestratorPlanner Agent
↓Dynamic routing
ContextRead materials
ResearchFind missing context
WriterCreate draft
Traceable workspaceShared State + Evidence
↓Writer Agent
Reviewer Agent
↶ Revise / gather evidenceSafety boundaryHuman Approval
↓Email ActionMocked in current MVP
The product decision
Why not just use a single prompt?
Different communication tasks need different workflows, context and verification depth.
Prompt → LLM → Email
- User must organize all context
- The same workflow handles every task
- Limited verification
- Weak handling of missing information
Goal → Plan → Adapt → Verify
- Planner determines what the task needs
- Only relevant agents are activated
- Claims are grounded in evidence
- Review can trigger re-planning
The agent architecture exists to make the workflow fit the task — not to add complexity for its own sake.
Evaluation
Measured, then improved.
An 80-case safety suite covers job, PhD, school and ambiguous requests. Results below separate deterministic workflow checks from a small LIVE model run.
Intent Accuracy
V2 regression result across the 80-case deterministic safety suite.
100%Task Completion
Contract completion, including correct ASK_USER behavior.
100%Hard-case Completion
Up from 41.0% before evaluation-led routing changes.
100%LIVE Qwen3 Completion
12-case ContextMail run vs 75.0% for the single-LLM baseline.
58.3%↗
The LIVE result did not yet justify the full agent cost.
On 12 stratified Qwen3 8B cases, ContextMail was slower (91.0s vs 16.8s), used 2.83 vs 1.0 calls, and completed fewer tasks. The next product focus is reducing reviewer/re-plan loops and validating with real evidence.
Delivery status
Live locally. Safe in the demo.
The static portfolio never calls external services; authenticated capabilities are available only in local LIVE mode.
✓Adaptive Planner routing
✓Brave Search provider
✓Traceable Evidence Store
✓Writer / Reviewer re-planning
✓Microsoft Graph Mail.Send
✓80-case evaluation suite
✓Human approval boundary
○Blinded human draft ratings
○Real-user edit distance
○Authenticated research trials
○Outlook delivery testing
○Cost / latency tuning