Skip to content
MOVAX Solutions

Work

Systems we built, and what changed because of them.

Real engagements, described without client names where we don't yet have permission to publish them. Ask on a call and we'll walk you through the specifics, architecture included.

Every one of them ships with

  • Langfuse tracing
  • Prompts versioned outside the code
  • Evaluation suites in CI
  • PII redaction before traces leave
  • Cost & latency budgets
  • Staged deploys & rollback

Professional services

Consultant matching agent

A text-to-SQL agent that reads a project brief in plain language, turns it into analytical queries across a consultant population of roughly 98,000 people, and returns a ranked shortlist with the evidence behind every match. Search strategy lives entirely in versioned prompts rather than in Python, so the model and the retrieval approach can change without a deploy. Generated SQL runs read-only against a cost ceiling.

Shortlists drawn from ~98,000 candidate profiles, with the reasoning attached

  • Agno
  • BigQuery
  • Python
  • Langfuse

Talent marketplaces

Semantic expert search

Retrieval over millions of experience facts, wired into an agent that can search, filter, rerank, and find lookalikes as a conversation rather than a search box. A query is expanded into several variations, each embedded and run as a kNN search, fused with reciprocal rank fusion, aggregated per expert, then reranked by a model against the original brief.

Every match traceable to the specific experience that produced it

  • Elasticsearch
  • Agno
  • OpenAI embeddings
  • Langfuse

Management consulting

Deep-research report generator

A consultant states a question, attaches whatever material they have, and gets back a sourced report plus whatever else the use case declares: PDF, DOCX, spreadsheet. Figures are computed arithmetically from operands that each trace to a quote rather than asserted by the model, and a critic strips any claim the evidence doesn't support before delivery. Report types are data, not code, so consultants author their own.

Every figure traceable to a quote, or removed before it ships

  • Agno
  • Langfuse
  • Elasticsearch
  • WeasyPrint

Sales operations

Sales call scoring pipeline

An ETL pipeline that classifies years of recorded calls against a fixed scoring rubric and turns them into per-rep performance metrics. The deterministic half of the rubric stays in code: counters, totals, and thresholds are arithmetic, and only the genuinely subjective judgments go to a model. Every score carries the passage it came from, so a manager audits a number instead of trusting it.

Every call scored on the same rubric, with the evidence attached

  • Databricks
  • Agno
  • Spark
  • Langfuse

Revenue operations

Account intelligence pipeline

Calls, CRM meetings, and message threads collected nightly, run through signal extraction, relevance filtering, and urgency scoring, and turned into recommended next actions on the accounts that need them. Built as a durable task chain with retries and ordering guarantees, because a stage that silently races another produces confident nonsense.

Signals surfaced with a recommended action, not another dashboard

  • Celery
  • Agno
  • Databricks
  • Langfuse

Sales enablement

Evidence-backed deck generator

A brief in, a client-ready deck out. Candidate engagements are found in the warehouse, reranked by a model on topical fit rather than deal size, deduplicated across four layers, and enriched with real outcomes. A model-as-judge critic then decomposes every slide into claims and attributes each to a source quote, so nothing reaches a prospect that isn't backed by a record.

Decks assembled from real engagement records, claim by claim

  • Python
  • MySQL
  • Agno
  • Langfuse

Talent platforms

Self-enriching onboarding

A signup flow that does the typing for you. An agent researches the person registering across public sources, drafts a complete profile with a citation behind each field, and hands it over for review before anything is saved. Nothing is written without the person confirming it, which is what makes an automated profile acceptable rather than creepy.

Profile creation turned from a long form into a review step

  • Next.js
  • Agno
  • Python
  • Langfuse

Enterprise IT

Internal agent builder

A platform where any employee can assemble an agent for their own work: choose the model, write and version the prompt, pick which tools it may reach, then share it with their team. The organisation keeps control of spend, access, and audit, so enabling everyone doesn't mean losing track of what is running or what it costs.

Agent building moved from the AI team to the people with the problem

  • Next.js
  • Agno
  • Langfuse
  • Postgres

Where we work

Different sectors, the same failure modes.

Messy documents, decisions buried in data, and processes that don't scale with headcount. The domain changes; the engineering problem rarely does.

  • Financial services

    Document-heavy onboarding, risk scoring, and reconciliation with a full audit trail.

  • Legal & professional

    Contract analysis, clause extraction, and grounded research assistants over case files.

  • Retail & e-commerce

    Demand forecasting, catalogue enrichment, personalised ranking, and support automation.

  • Healthcare

    Clinical document structuring, scheduling agents, and privacy-preserving on-premise deployment.

  • Logistics

    Route and demand prediction, proof-of-delivery vision, and exception handling at scale.

  • Manufacturing

    Visual quality inspection on the line and predictive maintenance from sensor history.

  • Education

    Course assistants grounded in real material, plus grading and feedback support tools.

  • SaaS & startups

    AI features inside an existing product, from first prototype to a metric you can defend.

Next step

Your project could be the next one here.

Tell us what you're trying to build and we'll walk you through the closest system we've shipped, architecture included.

Book a discovery callcontact@movaxsolutions.com

Cookies

We use local storage to remember this choice. Accepting also enables analytics, which help us see which pages are useful. Declining changes nothing else about the site.

PrivacyTerms