Soich Inc. — AI agents on your data

AI in production. On your data.Soich Inc. builds AI agents that read your data and carry out a task, with guardrails against invention: cited sources, a closed vocabulary, answers validated at runtime. The site you are reading is one of them — its agent can only cite verified content.Underneath, the engineering that makes them reliable: a layered Lakehouse on Databricks and Azure, model lifecycle with MLflow, self-hosted open-weight models when data must not leave. And at the far end, agents embedded on a robot — Reachy Mini, with persistent graph memory.

Our Services

What we build, and with which mechanisms — detail rather than promise.

AI & Intelligent Agents

Agents that look into your data and carry a task through — with guardrails that stop them inventing: answers validated at runtime, with a fallback. The site you are reading is one.

Document search with cited sources, over a graph (Neo4j) or vector store (Qdrant)
Closed vocabulary: nothing outside a verified catalogue
Self-hosted open-weight models when your data must not leave
LangGraph and n8n orchestration, tools called explicitly

Databricks Solutions

Databricks is not only a data platform: it is also where the models live. Ingestion, Lakehouse and governance on one side; MLflow, reinforcement learning, RAG and agents on the other.

Ingestion, layered Lakehouse and access governance
Model lifecycle with MLflow: tracking, registry, serving
Reinforcement learning and purpose-trained models
RAG and agents served from the platform, next to the data

Automation & Integration

Repetitive work moves to the background. Exceptions are handled explicitly — that is where automations usually break.

End-to-end processes orchestrated on n8n
Connection to existing systems through APIs
Written recovery paths for out-of-norm cases
Logging and failure alerts, no silent breakage

Robotics & Embodied AI

Agents that step off the screen. Already built and presented at conferences: Reachy Mini, persistent graph memory, connection to Databricks Genie.

AI agents embedded on Reachy Mini (Pollen Robotics), vision and perception
Persistent memory through a graph database (Neo4j)
Robot connected to an enterprise analytics platform
Demonstrators shown at user groups and AI Tinkerers

Analytics & Decision

A metric defined once and computed in one place — an end to twelve versions of the same number depending on the dashboard.

Metric definitions written down and versioned
Computed centrally, not copied into every report
Reports produced and distributed automatically
Self-service access on a controlled foundation

Certified Databricks Partner

What this status changes for your projects, concretely — and what it does not.

A direct line to the Databricks teams
Access to previews and betas, before release
Certification kept current, as the status requires
Regular talks at Databricks user groups
Attendance at the Databricks Summit
First-hand visibility on the roadmap

Official training and certification on the Databricks platform

Priority access to Databricks technical teams

Access to the latest features and best practices

Our Approach

A method that validates before committing, measures what it delivers, and bills only the hours actually spent.

A prototype before committing further The first iteration confirms technical feasibility and the quality of results on your real data. Nobody commits to six months blind.
Evaluation is a line in the quote Measuring the quality of what is produced is not an end-of-project intention: it is a budgeted item, like development itself.
An estimate, component by component Each building block is costed separately, with written assumptions. Infrastructure costs are kept apart from fees, so you know what you are paying for.
Billed on hours actually spent Billing follows the hours worked. If the objectives are met before the planned budget, the balance is not invoiced.

How a project unfolds

Framing The need, the available data, the usage constraints and the volume assumptions are written down before any estimate.
Prototype A short iteration validates the architecture and the quality of results on a real sample of your data.
Iterations Short cycles, each delivering a usable version. The cadence is set during framing, and development is accelerated by AI.
Evaluation & acceptance Thorough testing, measurement of answer quality, acceptance with your teams, fixes.
Progressive rollout Start with a handful of users, then widen by cohorts, so adoption keeps pace with the technology.

Curious by passion, cautious by experience

Every new thing is tried out on our own demonstrators before it is ever proposed to anyone — and most of them do not survive the test. That distance is what makes it possible to say no to a fashionable tool, and yes to the one that will still hold in three years. You get the benefit of the watching without paying for its mistakes.

Industries

Four sectors, each with the use cases handled and the technologies involved. The technical work itself is described under services.

Retail & E-commerce

Analytics solutions to optimize sales, manage inventory and improve customer experience.

Behavioural analytics
Inventory optimisation
Recommendations
Demand forecasting

Finance & Banking

Large-scale data platforms, regulatory reporting and risk analysis, with the traceability requirements the sector imposes.

Financial reporting
Fraud detection
Credit risk analysis
Compliance

Pharmaceutical & Healthcare

Data analysis solutions for research, production and compliance in the healthcare sector.

Clinical trials
Supply chain
FDA/EMA compliance
Pharmacovigilance

Manufacturing & Industry

Optimization of industrial processes through predictive analytics and IoT to improve productivity.

Predictive maintenance
Production optimisation
Quality control
Performance analysis

Talks & Demos

Presentations and demonstrations of AI & Robotics projects

Reachy Mini x Genie — Databricks

Demonstration of the integration between Reachy Mini and Databricks Genie, connecting a physical robot to an enterprise analytics platform.

Reachy Mini & Graph Memory

Presentation of Reachy Mini with persistent memory via Graph DB, enabling the robot to remember past interactions.

Ready to transform your data?

Let's discuss your project and discover how we can help you create value from your data.

infos@soichinc.com

Montreal, Quebec, Canada & Remote