01 / State
Persistent context
Conversation and workflow state retained across turns, channels and asynchronous events.
Production agent systems
Specialist AI agents, engineered for production.
SeedPath builds agents for defined operational work—complete with state, tool use, validation and structured outcomes. Deploy them through our API, or work with us to build and optimise a system for your organisation.
01Customer channel
↓
02SeedPath Agent API
↓
State + contextretained
↕
Specialist modeltara-v4
↕
Tools3 available
↓
05Validation
↓
06Structured outcome
Execution trace
thread th_91b8
intent book_appointment
tool outlook.find_slots
action schedule_appointment
validation passed
intent book_appointment
tool outlook.find_slots
action schedule_appointment
validation passed
Where to start
Start with what you need the system to do.
01 / The system
Four systems, one conversation.
Optimised to nurture, route and recover missed opportunities. From low intent nurture to live call transfer, we've developed our platform to catch as many opportunities as possible and ensure that they convert.
01 Lead sourceportal / partner / CRM
↓
02 Mya — Originationengages at low intent
↓
03 Tara — Engagementholds the conversation
↕
04Call-Bridge — Routing & recovery
↓
Automated call deliveryoutbound dial
↓
Missed-call recoveryrequeued → Tara
↓
Post-call analysissentiment + outcome
↓
Negative-turn recoveryre-engaged → Tara
↓
05 Outcomeconfirmed / handoff
Example thread
lead ld_4471
stage call-bridge.recovery
event missed_call → requeued
sentiment negative → neutral
outcome mortgage_confirmed ✓
stage call-bridge.recovery
event missed_call → requeued
sentiment negative → neutral
outcome mortgage_confirmed ✓
Call delivery
Scheduled calls placed and connected automatically, on time without a human dialler.
Missed-call recovery
A call that isn't answered isn't dropped — the lead is routed straight back to Tara to continue the conversation by message.
Post-call analysis
Every call is scored for sentiment and outcome, not just logged as completed or missed.
Negative-turn recovery
Where sentiment drops mid-conversation, the thread is flagged and re-engaged rather than left to go cold.
↩Separately, aggregated sentiment and outcome data across all conversations feed back into evaluation and the next model fine-tune — a slower cycle than the recovery loop above, but the one that improves Tara over time.
These are the measures we report on for every deployment. Figures are specific to each client's traffic and are shared directly as part of ongoing reporting, not published here.
02 / The runtime
The model is only one part of the system.
A production agent has to remember, decide, act and account for what it did. SeedPath provides the operating layer around the model, so customers do not have to assemble it themselves.
02 / Tools
Typed actions
Defined tools with controlled arguments, execution conditions and recorded results.
03 / Validation
Controls before action
Outputs checked against business rules and operating boundaries before they progress.
04 / Observability
An inspectable record
Inputs, routing decisions, tool calls, latency and outcomes available for review.
05 / Escalation
Humans where needed
Ambiguous, sensitive and exception cases routed through defined human-review paths.
06 / Evaluation
Behaviour measured
Scenario suites test workflow accuracy, consistency, compliance and regression.
07 / Models
Specialist inference
Fine-tuned production models for defined work, with model-agnostic fallback paths.
08 / Outcomes
Structured results
Responses, actions and state changes returned in forms downstream systems can use.
03 / Two routes
Use the system at the layer that suits you.
Deploy an existing specialist agent inside your infrastructure, or work with SeedPath to design and operate a system around a specific enterprise workflow.
Agent API / Private access
Bring the channel.
We provide the agent.
Send events or messages through a REST API. SeedPath manages the conversation state, specialist model, tool selection, validation and structured output.
- REST API and webhooks
- Persistent multi-turn state
- Customer-defined tool interfaces
- Structured actions and outcomes
Enterprise systems
Built around a defined operational problem.
We design, integrate, evaluate and continuously improve agent systems for organisations with specific workflows, data environments and control requirements.
- Workflow and failure-mode design
- System and CRM integration
- Evaluation suite development
- Fine-tuning and ongoing optimisation
04 / Model programme
We fine-tune models for the work they actually perform.
General-purpose models are designed to answer almost anything. Our production agents operate within narrower environments, with defined actions, constraints and measures of success.
A small, targeted change can produce a material behavioural gain.
Tara is a SeedPath LoRA fine-tune of Llama 3.1 8B Instruct. Its validators-first output structure makes validation part of the context used to generate the response and select an action—not a filter added afterwards.
We describe the foundation model, training method, holdout boundary, evaluation method and known limitations. “Proprietary AI” is not a technical explanation.
Read the Tara model note →Schedule-action match52% → 91%
Adapted matrices224
Network-wide relative shift2.68%
Update energy in top 4 directions73%
INTERPRETATION — A targeted behavioural improvement on the defined mortgage-engagement task. This is not a claim of general model superiority. Evaluation scope and methodology should be read with the result.
05 / Evaluation
We evaluate behaviours, not vibes.
A fluent demo is not evidence of a dependable agent. Releases are tested against the decisions, actions, constraints and failure modes that matter in the workflow.
| Behaviour | How it is tested |
|---|---|
| Action selection | Expected workflow action match |
| Conversation quality | Scenario-based assessment |
| Compliance | Required and prohibited behaviours |
| Tool use | Tool, arguments and execution conditions |
| Escalation | Detection and correct human routing |
| Regression | Repeatable suite before release |
06 / Production controls
Security is part of the architecture.
Tenant separation, controlled data retrieval, authenticated ingress and human escalation are system properties—not badges added at the end of the page.
Isolation
Tenant-scoped data
Customer identifiers and database row-level controls prevent cross-tenant retrieval.
Ingress
Verified requests
Authenticated HTTPS endpoints and signature verification before processing.
Context
Controlled retrieval
The orchestration layer retrieves required context; retrieval is not delegated to the model.
Oversight
Defined review paths
Negative, ambiguous and exception cases can be stopped and routed for human review.
07 / Technical notes
Our R&D work, including the awkward parts.
Research notes, model cards and architecture decisions, provided here to support better decision making on AI products.
What changed inside Tara during fine-tuning
Weight-space analysis of 224 adapted matrices, and what it suggests about learned routing and decision behaviour.
Read the note →
Why we don't let the model make the decision
Five ways of asking an AI to decide something, tested across 50,000 conversations — and why the one that works doesn't ask it to decide at all.
Read the note →
Where we still use frontier models—and why
A practical account of primary inference, lightweight classification and model-agnostic fallback.
Start with the work the agent needs to complete.
Tell us about the workflow, its systems, its constraints and what a successful outcome looks like.
Discuss an enterprise system →