01 / Custom LLM software

We build software where language models do the work, not demonstrate it.

Custom development for document-heavy business processes: structured data extraction, integrations with external APIs, internal operational tools, and AI agents that run under rule-based control.

Start a conversation info@corpus-ai.am

Fig. 01 — Document processing pipeline
  1. 01 IntakePDF · scan · email · API
  2. 02 Extraction & classificationLLM pass → typed schema
  3. 03 Rule validationdeterministic checks, human review on conflict
  4. 04 System of recordERP · CRM · warehouse
02 / SERVICES

Four directions of engineering practice.

01

Document processing automation

Extraction of structured data from contracts, invoices and forms; classification, cross-field validation and routing into the systems that already run the business.

  • OCR
  • LLM extraction
  • JSON schema
  • Human-in-the-loop
02

Integrations and backend

Connecting internal systems to external APIs: queues, idempotent jobs, retries and explicit error handling, so that a failed request is a recorded event rather than lost data.

  • REST
  • Webhooks
  • Queues
  • PostgreSQL
03

AI agents for operations

Scenarios where a model performs a multi-step task under rule-based control: bounded tool access, deterministic checkpoints and a full audit trail of every action.

  • Tool use
  • Orchestration
  • Guardrails
  • Evaluation
04

Internal web applications

Admin panels, CRM modules and operator tools built around the real workflow of the team that uses them, with role-based access and reporting.

  • TypeScript
  • React
  • Auth & roles
  • Reporting
03 / PROCESS
  1. 01 Discovery and specification Process review, data samples, written scope with acceptance criteria.
  2. 02 Prototype in 1–2 weeks A working slice on real documents, so feasibility is proven before the main build.
  3. 03 Iterative development Sprints with demo builds, a tracked backlog and versioned releases.
  4. 04 Handover and support Documentation, deployment instructions, code transferred to the client's repository.

Stages are fixed and priced separately. Each stage closes with a signed acceptance act. Source code, infrastructure configuration and technical documentation are transferred to the client's own repository and accounts — there is no vendor lock-in on our side.

04 / STACK
Languages & runtimes
  • Python
  • TypeScript / Node.js
  • SQL
  • Bash
Models & AI infrastructure
  • Claude API
  • OpenAI API
  • open-weight models (self-hosted)
  • embeddings & vector search
  • structured output / tool use
  • evaluation harnesses
Data & storage
  • PostgreSQL
  • Redis
  • S3-compatible object storage
  • pgvector
  • ETL pipelines
Infrastructure & deployment
  • Docker
  • AWS
  • Hetzner
  • GitHub Actions
  • Nginx
  • observability & logging
Development tooling
  • Git
  • pytest
  • type checking
  • OpenAPI
  • code review
  • staging environments
05 / APPLICATIONSanonymised engagements
01

Incoming invoice processing for a distribution business

TASKSeveral hundred supplier invoices per week arriving as PDFs and scans, entered into the accounting system by hand.
APPROACHExtraction into a typed schema, matching against the purchase order, automatic flagging of discrepancies for an operator.
RESULTManual entry replaced by review of exceptions only; every document keeps a traceable link between source file and record.
02

Contract review support for a legal function

TASKRecurring review of supplier agreements against an internal checklist of required and prohibited clauses.
APPROACHClause segmentation, retrieval against the internal policy set, a per-clause report with citations to the source paragraph.
RESULTLawyers work from a structured report instead of a blank document; conclusions remain human, evidence is machine-collected.
03

Operational agent over internal systems

TASKRoutine multi-step requests handled by an operations team across three disconnected internal tools.
APPROACHAn agent with a bounded set of tools, rule checks before each write action and a full log of decisions.
RESULTRoutine requests are completed unattended; anything outside the rules is escalated with the reasoning attached.
06 / COMPANY

CORPUS AI LLC is a software development company registered in Yerevan, Republic of Armenia. The practice is narrow by design: applied engineering around language models for document-heavy and operational processes, delivered as custom software rather than a product licence. We work directly with the client's technical and operations staff, in English, Armenian and Russian.

Legal name
CORPUS AI LLC
TIN
08329676
Jurisdiction
Yerevan, Republic of Armenia
Domain
corpus-ai.am
Working languages
English · Armenian · Russian
07 / CONTACT

Write to us with a short description of the task — the process, the document types involved and current volumes are enough to start.

Start a conversation

Opens your mail client with a message to info@corpus-ai.am. Technical and commercial enquiries are answered within one business day.