Miranium AI / Services

Applied AI, built and
evaluated.

Five service lines for organizations whose AI gets audited. We build AI systems for document-heavy, evidence-bound processes, evaluate and benchmark the AI you already run, and keep the record of both. Every engagement is delivered by the principals.

Engagement
Direct · Subcontract · Teaming
Delivery
Founder-led · delivered by our own engineers
Place of performance
Remote · on-site as required
What you get

What each line actually delivers.

The same five the home page introduces, in statement-of-work terms. Each is one posture applied to a different problem:

  • Get the evidence
  • Compute what can be computed
  • Let a model narrate rather than decide
  • Leave a record a reviewer can follow

Most engagements draw on two or three at once.

01

Document Intelligence & Process Automation

Structured data out of unstructured and semi-structured source documents: invoices, forms, claims, scanned records, validated against your business rules and written to your system of record. Deterministic OCR first, then a governed model pass over the result: the model reviews the result, and deterministic code writes the number into your ledger.

Scoped for a state port authority: AP invoice automation across four capture channels, three legal entities, and PeopleSoft FSCM 9.2.

  • Two-stage pipeline: OCR engine, then governed model validation
  • Every field carries a confidence score and a pointer to the region it was read from
  • Sub-threshold fields route to a review queue rather than posting
  • Exception workflow with SLA timers, severity, and escalation
  • ERP write-back, web service, component interface, or staged file
  • Immutable audit log of every extraction, coding, and approval
02

AI Governance & Assurance

The controls that let an organization put an AI system in front of a regulator. Human-in-the-loop architecture where every model-proposed action is logged as a reviewable proposal, autonomy is configurable per class of action, and the record of what was proposed, who approved it, and what ran survives the review afterward.

  • Approval contract: propose → approve → execute, one durable ledger
  • Autonomy granted per reversibility class
  • Control mapping and assessment against the NIST AI RMF
  • Model evaluation harnesses, with published failure modes and thresholds
  • Production monitoring for drift, hallucination, and silent degradation
  • Independent review of an existing AI system, including a vendor’s
03

Applied AI Systems & Agent Engineering

Custom AI systems built the way ours are: a tool-calling agent over your own data and your own systems, with deterministic guardrails wherever the output becomes a figure, a decision, or a record. Grounded retrieval so answers cite a source passage, and abstention so the system declines rather than guessing well.

  • Hybrid lexical and semantic retrieval over your corpus
  • Multimodal intake, photo, PDF, scan, or plain text
  • Deterministic engines behind anything that becomes money
  • Tenant isolation and capability-scoped access, per person
  • Per-turn telemetry: steps, guardrails, latency, and token cost
  • Evaluation suites that run offline, so behavior is regression-tested
04

AI Evaluation & Benchmarking

Evidence that an AI system does what it was built to do, measured rather than asserted. Your agents run through test suites scored by meaning across nine quality dimensions, every cited source is verified, and each release is compared with the last. Every score carries a written explanation, so a reviewer can follow it from the number back to the response.

Runs on MiraEval and MiraBench, the two evaluation platforms we built, shown below.

  • Single-prompt, multi-turn, and agentic task tests against your own agents
  • Nine scored quality dimensions, each with a written explanation
  • Grounding, policy, and citation checks against your knowledge sources
  • Baselines, release comparison, and quality gates in your CI/CD
  • Load and latency testing, then continuous production monitoring
  • Behavioural benchmarks of language models in markets and auctions
05

Data Engineering, Analytics & Decision Support

Turning the output of everything above into something a person decides on. Extraction and classification over unstructured data, cost and risk estimation engines, and dashboards that report to operators and to leadership with a single version of the number.

  • Pipelines from source documents to a queryable, versioned record
  • Deterministic cost and risk estimation engines
  • Operational dashboards: throughput, cycle time, exception aging
  • Executive reporting, and a documented API or export for your BI tool
  • Cohort and funnel analysis over behavioral data
  • Metric definitions written down, so a number means one thing
AI evaluation, in practice

Two evaluation platforms we built.

Our evaluation engagements run on software we wrote and operate ourselves. MiraEval tests the AI agents an organization deploys; MiraBench measures how language models behave when they make decisions. Both come as part of an AI evaluation engagement.

MiraEval · AI agent testing and evaluation

MiraEval tests, scores, and monitors conversational chatbots, virtual assistants, and task-based agents across development, test, and production. Every response is scored by meaning across nine quality dimensions, each with its explanation; cited sources are verified, releases are compared side by side, and results land in dashboards and in PDF, Excel, and JUnit reports for QA teams, product owners, and leadership.

  • REST, OpenAI-compatible, Claude, Microsoft Teams, and browser chat connectors
  • Test cases generated from your knowledge sources, reviewed before use
  • Quality gates in GitHub Actions, Azure DevOps, Jenkins, and GitLab
  • ServiceNow, OpenText ALM, and Jira integration
  • Load testing with virtual users, latency percentiles, and throughput
  • Okta SSO, role-based access, encrypted records, hash-chained audit log
The MiraEval dashboard for a demonstration project, SRS Member Services Assistant: pass rate 87.5 percent, average score 86.1, quality gate met with 7 of 8 cases passed, and p95 latency of 1075 ms, above a quality trend across thirteen runs, a response classification chart, bars for nine quality dimensions, and a list of alerts
The project dashboard: pass rate, average score, quality gate and latency, with quality tracked across releases and every score one click from the prompt that produced it.
A single MiraEval result: a retiree asks whether they can go back to work, the agent answers, and the expected answer sits beneath it. Beside it the evaluation scores 94 with a pass, broken down by semantic accuracy, completeness, relevance, clarity, consistency, grounding, policy alignment, citations and task completion, each with a one-line explanation, and a human review panel to confirm the verdict
One response, scored. Each dimension states its reason in plain language, and a reviewer can confirm or adjust the verdict beside it.

MiraBench · behavioural benchmarks for language models

MiraBench places language models in markets and auctions where every value is set in advance, so the efficient outcome and the rational reading of every news item are known exactly. Each model’s choices are then measured against rational, algorithmic, and human reference traders: price formation, anchoring, herding, risk preference, response to news, and how an AI advisor shapes a person’s decisions. It works through text alone, so any model can be benchmarked, from frontier APIs to open-weight models run locally.

  • Six market mechanisms: double auction, call market, sealed-bid, English, and Dutch
  • Anthropic, OpenAI, Gemini, Mistral, DeepSeek, and local open-weight models
  • An exact Bayesian benchmark for every news item
  • A human-vs-AI classifier and a decision-level comparison with human proxies
  • Leaderboards with paired, Holm-corrected tests across matched seeds
  • Web UI with a live market view, REST and WebSocket API, and a CLI
A MiraBench run of eleven language models trading in a continuous double auction over three periods: allocative efficiency 95.9 percent from 29 trades and 512 decisions, a market price chart of best bid, best ask and trade prices against the Bayesian fundamental and the competitive equilibrium range, efficiency by period at 96, 96 and 95 percent, and a table of each period
Eleven models trading in one market. Prices are plotted against the Bayesian fundamental and the competitive equilibrium, with efficiency measured period by period.
The MiraBench agents table for the same run: eleven language model agents, four human-proxy agents, two algorithmic traders and two held-out test agents, each with its family, role, profit, decisions, distance from human decisions, probability of being human, and behavioural flags such as round numbers, anchoring and momentum
Every agent in the same run, side by side: profit, distance from human decisions, and the behavioural patterns each one shows.
How we deliver

Discovery is where the price gets honest.

A phased engagement with a decision point at the end of each phase. The scope is settled in the first weeks, and the estimate you sign is confirmed against your real volumes before any build starts.

6phases, each ending in a decision you make
16–20weeks from discovery to cut-over
60days of stabilization with daily monitoring and named exit criteria
5business days’ notice before any change to an integration, workflow or schema
  1. Discovery & gap analysis

    Stakeholder workshops, current-state documentation, and a future-state design you review before anything is configured. Accuracy targets and success metrics are agreed here, in writing.

    2–3 weeks
  2. Design & configuration

    Rules, routing, tolerances, and integration points designed against that future state, since translating today’s manual process into software is how you automate a workaround.

    3–4 weeks
  3. Build & integrate

    Built in non-production, in code, under version control. Separate development, testing, and production environments, provisioned identically, with reviewed promotion between them.

    5–6 weeks
  4. Test & validate

    User acceptance on your real, de-identified data. Accuracy tuning against the targets set in Discovery, then a go/no-go readiness review that you chair.

    3–4 weeks
  5. Train, cut over, stabilize

    Role-based training with recorded sessions, written SOPs specific to your configuration, then a stabilization period with daily monitoring and named exit criteria rather than an open-ended tail.

    3 weeks + 60 days
  6. Operate

    Support against the service levels below, a published release cadence, and a review cycle that retunes thresholds against what the last quarter actually did.

    Ongoing

Every change reaches production through your review

Anything touching an integration, an approval workflow or a data schema is communicated five business days ahead and needs your approval before it deploys. Backward-compatible updates ship on a published cadence with notice. Every production change goes through the pipeline, since that is where it gets tested.

Service levels

What we commit to in writing.

Baseline targets for a managed engagement. They are negotiable upward and they go in the contract, because a service level becomes a commitment only when it is signed.

Baseline service level targets by severity
P1, Critical System down; the process is halted. 15-minute response, 24×7, continuous effort until resolved or a workaround is in place.
P2, High A major function impaired, such as a failing integration. 1-hour response, resolution targeted within one business day.
P3, Medium Non-blocking defect or degraded accuracy on a subset. One business day to respond, five to resolve.
P4, Low Cosmetic, question, or enhancement. Two business days to respond; scheduled into the next release.
Availability 99.9% monthly for hosted services, measured and reported, with maintenance windows announced 72 hours ahead.
Reporting Monthly by the tenth business day, availability, incidents by priority, backups verified, and the metrics that matter to your process. Quarterly strategic review.
Support hours 24×7 on-call for P1. P2–P4 Monday to Friday, 8:00 AM–6:00 PM ET.
How to buy it

Four ways we contract.

Small business, SAM.gov registered, Virginia SWaM certified. We prime engagements sized for a two-person senior team and subcontract on larger scope, and we tell you which before you bid. The first increment is sized to fit inside a quarter, baselined before we build.

Fixed-price, phased

A firm price per phase against a scope agreed in Discovery, the total confirmed against your real volumes before design or build begins. Year-one hosting, processing and support are separate lines, so you can see what recurs.

  • Priced per phase, with a go/no-go between
  • Recurring costs itemized up front

Managed-service retainer

A monthly or annual figure for operating something continuously: hosting, monitoring, scheduled evaluation runs, support against the service levels above, and the reporting that proves it. Multi-year terms with an annual re-confirmation against actual volume.

  • Service levels in the contract, reported monthly
  • Re-confirmed against real volume each year

Time and materials

A four-role rate card, project lead and solution architect, data and AI lead, integration engineer, QA and validation, for change orders, out-of-scope work and advisory that suits a flexible scope.

  • Published rates, senior-weighted
  • Specialist support disclosed and billed explicitly

Subcontract & teaming

As a small-business subcontractor to a prime, or teamed on a bid where we bring the AI, data, or platform scope. NDAs, MSAs, and teaming agreements welcome. SBIR and STTR paths considered.

  • Small · micro · women-owned · minority-owned, Virginia SWaM #828797
  • UEI D9WZWV4JMR33 · CAGE 9KE77
What we work in

The stack we actually operate.

Only tools our principals have run in production, under load, with a pager attached.

AI & machine learning

Tool-calling agentsHybrid retrieval Vector searchOCR & document AI Multimodal intakeClaude OpenAIGemini Open-weight modelsPython

Evaluation & benchmarking

MiraEvalMiraBench Semantic scoringLLM judges Regression gatesLoad testing PlaywrightStatistical testing

Data & platform

PostgreSQLSQLite AWSGoogle Cloud GovCloudKubernetes TerraformGitHub Actions

Frameworks we build to

NIST AI RMF 1.0NIST SP 800-53 Rev 5 SOC 2 TSCHIPAA Security & Privacy Section 508WCAG 2.1 AA