Skip to Content
pyxl
Build a PrototypeStart a Conversation
Contextual · PartnerB2B

The expertise is in three people’s heads.There is a queue forming behind them.

Contextual is the production AI layer for judgment that currently lives in a few people's heads. It captures how experts reason, turns that reasoning into a system, and keeps a human in the loop when the system is uncertain.

Where the judgment lives
A task only two or three people can do
CapturedStructured intake gathers what a specialist would ask for
EncodedA rules engine holds how they reason, not just what they answered
AppliedEvery request gets the same expertise, in seconds
ImprovedEdge cases are reviewed by a person and sharpen the rules
90%
Reduction in valuation time, vintage instruments
<5%
Material error rate on pricing
400%+
Efficiency gain
8 wks
Rebuild to live use
The artifact

Ninety percent faster valuations,the same expert judgment.

Vintage guitar prices turn on condition, materials, damage and what the market did last month. The existing tool was slow and inconsistent, so the shop queued every final price behind a handful of specialists. This is what changed.

Before
After
Up to 60s
per valuation
90%
reduction in valuation time
Specialists
required for final price
<5%
material error rate on pricing
Inconsistent
results between staff
400%+
efficiency gain
Friction
in daily workflow
Retained
expertise, not tied to individuals
01Ask

A streaming conversational assistant gathers the right details, asking clarifying questions automatically rather than presenting a form.

02Evaluate

Structured inputs drive a rules engine that encodes how the shop's best appraisers reason about condition, materials and damage.

03Price

Historic sales blended with live market signal, so the number reflects today rather than the last time someone updated a sheet.

04Learn

Edge cases are reviewed by a human and fed back, so the rules sharpen instead of drifting.

Rebuild to live use, eight weeks
Week 1–2

Map the existing logic and capture the expert rules

Week 3–7

Build and test against real instruments

Week 8

Launch

Week 8+

Roll out to staff and tune on live usage

What we deliver

Four deliverables,from first build to steady state.

01

Expert systems, not chatbots

  • Structured intake with automatic clarifying questions
  • Rules engines that encode how your specialists reason
  • Retrieval over your own catalog, pricing and history
  • Human review on the edge cases, feeding the rules back
An expert system, end to end
AskStreaming conversation with automatic clarifying questions
EvaluateStructured inputs into a rules engine
Price or decideHistoric data blended with live signal
LearnHuman review on edge cases, fed back into the rules
02

Workflow applications

  • Two-sided portals over an existing system of record
  • Negotiation and approval lifecycles, not single transactions
  • Role-based access where each role is a different application
  • GraphQL layers so nothing reconciles twice
A workflow application, layer by layer
01System of recordSalesforce or HubSpot stays authoritative
02GraphQL layerOne contract, so nothing reconciles twice
03LifecycleAccept, reject and counter re-enter the same loop
04RolesOwner, admin and technician see different applications
03

Production discipline

  • Observability and evaluations before launch, not after a complaint
  • Guardrails on the outputs that touch money
  • Versioned prompts and rules, reviewable like any other code
  • Fallbacks when the model is uncertain, escalating to a person
Demo versus production
An AI demo
  • Impresses in the meeting
  • No evaluation set
  • Fails silently
  • One prompt, unversioned
Never reaches daily use.
A production system
  • Evaluations before launch
  • Observability on every call
  • Guardrails on money-touching output
  • Escalates to a person when uncertain
Runs every day and can be answered for.
04

How it is actually assembled

  • A router in front of the frontier models, so a deprecation is a config change rather than a rebuild
  • Retrieval over your own corpus with the chunking tuned to your document shape, not a default
  • A cache keyed on the normalised question, because the same question arrives a hundred times a day
  • Cost and latency recorded per call and attributed to a feature, so the bill is legible
  • Evaluations run against a fixed question set on every prompt change, in CI
The stack, layer by layer
01RouterSits in front of the frontier models; a deprecation is a config change
02RetrievalChunking tuned to your document shape, not a default
03CacheKeyed on the normalised question
04ObservabilityCost and latency per call, evaluations run in CI
05

Integration into the business

  • Salesforce, HubSpot and ERP as the system of record
  • Live market and pricing signal, not a stale sheet
  • Staff rollout and tuning on real usage
  • Knowledge retained in the system rather than in one person
What it has to talk to
The businessThe systemCRM recordCustomer and historyPricing dataHistoric salesLive marketCurrent signalStaff reviewRule refinement
Behind it

Delivered,and running today.

Retail & marketplace

The North American Guitar

90% faster valuations

A valuation assistant that prices vintage instruments in seconds, on condition, materials, damage and live market signal. Contextual published the case study; we did not.

Industrial & professional services

CASE Facilities Management

Procurement as a lifecycle

A two-sided service-partner portal over Salesforce, carrying a full bid, counter-offer and award lifecycle for a national snow and landscaping operator.

Software & AI products

Aivin

Concept to launch

An AI sommelier taken from concept to launch as a product, with conversational UX, brand identity and the orchestration underneath.

The valuation assistant answers in seconds what used to take a specialist twenty minutes, and it shows the comparable instruments it reasoned from, so a human can overrule it and the override trains the rules. That auditability is the part that took the eight weeks.
How the system works · The North American Guitar
Who this fits

Four situations
where this is the right tool.

The common thread is judgment that currently lives in two or three heads, and a queue forming behind them that hiring cannot clear fast enough.

01

A pricing, quoting or estimating task that only two or three people can do

02

A workflow already running in spreadsheets, email and a legacy portal

03

Expertise walking out the door with a retirement or a resignation

04

An AI pilot that impressed everyone once and never reached production

Free, and specific to ContextualSee what your Contextual instance is actually doingA senior strategist reads the account rather than running a scan: what is configured, what is actually running, and what is quietly costing you. You keep the findings whether or not anything follows.Request a Baseline

How these systems get used.One piece a month.

Written out of engagements rather than off a content calendar, by whichever principal ran the work. No sequence afterwards.

Your address goes into our own CRM and nowhere else. Privacy policy.

Is there a queuebehind one person’s judgment?

Tell us the task and the person everyone waits on. A senior engineer will tell you whether it is a rules problem, a retrieval problem, or not an AI problem at all.

Common questions

Questions we getbefore the first call.

What is Contextual?

Contextual is one of Pyxl's orchestration platforms for production AI — retrieval over your first-party data, routing across frontier models through their APIs, caching, and observability with cost and latency measured per call. It runs on your infrastructure. You own the application layer, the data and the corpus, and if our relationship ends you keep all of it and it keeps working.

Can you build a custom AI assistant for our business?

Yes. We have built nine systems now, including SONA for merchant applicant submission and approval at Celero Commerce, an AI sommelier for Aivin, a valuation and pricing agent for Carter Vintage Guitars, an ROI calculator wired into HubSpot for Vodori, a sales assistant for Relat and retrieval search across a one-of-one inventory for The North American Guitar. Each one is retrieval over the client's own data rather than a general model with a clever prompt.

Will we own the system, or are we renting it?

You own it. It runs on your infrastructure and calls frontier models through their APIs, so the application, your data and the retrieval corpus are yours. Subscription tools run on the same models everyone else uses and keep your data on their servers, which is the difference that matters in eighteen months.

Which AI models do you use?

Whichever performs best for the task, drawn through their APIs and routed per call. Contextual, our orchestration layer, handles routing, retrieval, caching and observability with cost and latency measured on every request, so a better or cheaper model can be swapped in without rebuilding anything.

How do you stop an AI system from saying something wrong to a customer?

A named person on your team approves anything generated before it publishes, and the gate is built into the product rather than written into a policy. Beyond that, retrieval over your own approved data narrows what the system can say in the first place, and we measure citation accuracy as a metric rather than assuming it.

We would like to measure which pages people read, using Google Analytics. That is the only thing these cookies do here — no advertising, no profile sold to anyone, nothing shared outside Pyxl. Say no and the site works exactly the same. Privacy policy.