MetisMesh is an AI-first software consulting firm. We design, build and operate the engines, platforms and assistants a business depends on, using intelligence where it earns its place and keeping everything correctness-critical deterministic.
One guiding principle: a commitment to excellence.
Six products we designed, built and run ourselves. Each one is a working answer to a difficult problem, and each one shows where we let intelligence in and where we keep it out.
What we are brought in to solve
The pattern is the same every time: understand the business rule, decide what must be certain, design the system around that, then build and run it.
A workbook nobody dares touch is the real system of record. We turn it into a deterministic engine with tests against the original numbers, so the migration is provable before it is trusted.
Scanned PDFs, exports, engineering drawings and policies with no way to ask a question. We build ingestion with provenance, retrieval that favors the right evidence, and answers that cite or abstain.
Tenant isolation enforced in the database as well as the code, consent and authorization as records, and audit trails that cannot be edited.
Rules engines, ledgers and governance cores where the answer is a pure function of the inputs and every decision can be replayed and explained later.
We draw the line: intelligence where retrieval, language, classification, summarization or automation creates value; deterministic code for calculation, authorization, transactions and audit.
How we think
Einstein's line is not a slogan here, it is a division of labor. Knowledge tells you what already exists. Imagination lets you see what could exist, question the assumption everyone stopped noticing, and design a better answer. Engineering excellence needs both, in that order.
And like a bicycle, a technology organization stays balanced by moving. Not chasing every trend, but continuously learning, modernizing, simplifying and automating, on purpose.
01
Imagination
See what could exist, not only what does.
02
Engineering judgment
Decide what must be certain and what may be intelligent.
03
Execution
Build it, test it against reality, run it in production.
04
Continuous forward motion
Learn, simplify, modernize and automate, deliberately.
05
Excellence
The standard the other four are held to, every time.
Selected work
Three of the six. Each story follows the same shape: the problem, what we built, where AI fits, the engineering discipline behind it, and what it is worth to the business.
A deterministic engine replaced a 130-sheet workbook. AI explains the numbers; it never touches them.
Kept deterministic
Demand and inventory calculation, Baseline approval and versioning, Tenant isolation and access, Cycle audit trail
Where intelligence helps
Study narrative from recorded data, Owner questions over org data, Concept explanations
The assistant can only see what a member has already chosen to share. That rule is proven by tests, not policy.
Kept deterministic
Disclosure and visibility, Consent and authorization, Introduction state machine, Verification and activation
Where intelligence helps
Conversation guidance, on request, Profile writing help, Operator research with citations
No verified progress, no capital progression. Every decision is a replay of an append-only, hash-chained ledger.
Kept deterministic
Tranche eligibility, Hash-chained ledger, Segregation of duties, Holds and certificates
Where intelligence helps
Findings that can raise an exception
Capabilities
We stay close to both the design and the code: from the business problem through data modeling, API and security boundaries, implementation, testing, deployment, observability and production support.
Planning engines, governance rules, ledgers and workflows where the answer must be the same tomorrow as today. We keep these deterministic, versioned, tested against reference data, and auditable by replay.
Python · Go · TypeScript · PostgreSQL · event sourcing · state machines
Document ingestion with provenance and OCR, chunking that follows structure, hybrid lexical and semantic search, cited answers, and abstention when the corpus cannot support a claim.
LangGraph · LangChain · Amazon Bedrock · OpenSearch · Vertex AI · AlloyDB · pgvector
Assistants that summarize, classify, screen, draft and explain inside workflows a person still controls. Budgets, circuit breakers, privacy filters and typed handoffs are part of the design, not an afterthought.
Claude · Bedrock · Vertex AI · LiteLLM · Amazon Lex · typed tool contracts
Tenant isolation enforced at the database and the application, consent and authorization as first-class records, append-only audit, and identity that fails closed.
Supabase · row-level security · JWT and JWKS · Better-Auth · OAuth2 · mTLS
Moving legacy estates to managed cloud without losing the business behavior that made them valuable: serverless edges, container platforms, messaging, infrastructure as code, and the regulated-environment discipline that comes with federal and financial work.
AWS · Azure · GCP · ECS and Fargate · Lambda · Terraform · CloudFormation · Kubernetes
Metadata-driven integration layers, near-real-time ingestion, telemetry that shows latency, error trends and backpressure, CI/CD, runbooks, and production support through the first incident and beyond.
Redis · RabbitMQ · Kafka-style pipelines · OpenTelemetry-style tracing · CI/CD · Docker
Languages, clouds and platforms we work in
Why clients trust the engineering
Federal programs under FedRAMP, FISMA and NIST controls, financial platforms with mTLS and least-privilege identity, healthcare integrations, and treasury knowledge systems. Security boundaries and repeatable deployments are normal engineering here, not a compliance add-on.
Six of the systems on this site are ours to run: users, incidents, migrations, cost. That experience is in every architecture we propose for you.
Authorization, calculations, transactions, audit and security boundaries are code and database constraints. Intelligence proposes; rules decide. Missing data stays missing instead of becoming an invented value.
Runbooks, walkthroughs and clean handoffs are part of the work. We stay through the first incident, then step back. Enablement, not dependency.
Code and database constraints. Same inputs, same answer, replayable later.
Assistance inside workflows people still control, with budgets, breakers and privacy filters.
The first conversation is thirty minutes with the person who would do the work. No sales script, no slide deck. Bring the constraint that is keeping you up at night.