Agent systems for high-stakes operations.

SKOPAQ AI PRIVATE LIMITED builds agent systems for operations where being wrong is expensive — designed, coded and gated by an autonomous agent workforce. Its first product, VELA, investigates satellite anomalies: it turns a telemetry deviation into a root-cause account and a drafted recovery procedure, every claim carrying its source, and a person executes every step. The long version — the products, the method, and what we refuse to claim — is at skopaq.com.

SKOPAQ AI PRIVATE LIMITED · incorporated 30 July 2026 · CIN U62011TS2026PTC220462 · Hyderabad, Telangana, India · skopaq.com · solutions@skopaq.com

Three AI systems, one of them open source. Each entry says what it is, what it is not, and whether you can read the source.

Live · private repository · case study

Reads a telemetry deviation against the spacecraft's own documentation and returns a root-cause account with a drafted recovery procedure. Advisory only: nothing it produces commands a spacecraft, and an operator executes every step.

• Every claim carries its source document. An uncited grounded step fails the CLI with a non-zero exit, and a release with an ungrounded claim fails the citation check in CI.
• A guardrail script greps for telecommand-issuance identifiers in pre-commit and in CI, so the advisory-only boundary is enforced by tooling rather than by intent.
• Deterministic: identical inputs produce an identical investigation. Every step appends to an audit log that is never rewritten, and a run can be replayed from its inputs and compared against a golden output.
• All spacecraft data is a synthetic reference graph plus seeded replay scenarios. None of it is presented as a real mission.
• Runs in the cloud or fully air-gapped from one configuration flag; CI exercises both.
• Aimed at the investigation gap the ESA-ADB benchmark identified. Not yet run against that benchmark: no score exists, and none is claimed.

Stack: Python 3.12, LangGraph, MCP, Pydantic v2, Neo4j and Qdrant, FastAPI, a Next.js operator console.

Public source, MIT · case study

Point it at a codebase and a running application. It reads the codebase, plans the coverage, drives the browser, checks the API and the database, and repairs the selector when the interface moves instead of failing the build: URL → discover → generate → export → execute → heal → report.

• Crawls with Playwright and vision models to find routes, forms and flows; exports to Playwright, Pytest, Jest, Cypress and Selenium; executes UI and API tests, capturing screenshots and traces.
• Model choice is configuration, not architecture: Anthropic, OpenAI, Azure OpenAI, AWS Bedrock, Vertex AI, OpenRouter and local Ollama are all supported. Claude is the default.
• The repository carries a 95% backend coverage badge. That figure is the coverage of Argus's own test suite, not a measure of how accurately it will test your application.
• A self-healing test that heals wrongly is worse than a failing test: a real regression turning into a green build is the silent failure the healer has to avoid to be worth running.

Repositories: argus-backend (the engine, MIT) · argus (the Next.js dashboard) · argus-mcp-server (MCP server for AI IDEs) · argus-plugin (Claude Code plugin, MIT) · skopaq (also carries the self-healing engine).

Argus also runs on skopaq.com: POST https://skopaq.com/api/argus-scan classifies a prompt for injection on Cloudflare Workers AI and returns the model's own verdict, technique and reasoning, the model id and a measured latency. There is no confidence score, because the model returns none.

Live at routera.dev · private repository · case study

An OpenAI-compatible gateway in front of every model you use. Routes requests across Google Vertex AI, Anthropic, OpenAI, DeepSeek and OpenRouter on your own provider keys at zero markup, with capability-aware routing and fallbacks, per-key budget circuit breakers that stop a runaway agent loop, and a live pricing matrix.

Published on skopaq.com and linked from every page, but not in the product catalogue.

• Pragya OS — offline-first school management for Indian government schools: three Next.js portals (admin, teacher, parent) over six services in a Turborepo monorepo. Built and not yet deployed. There is no portal URL and no school is using it. Private repository. Renamed from Shiksha OS on 21 August 2026. Case study.
• VERS India — a membership, governance and conference platform built and delivered for the Veterinary Education and Research Society, India. Client work, named with their consent. Live at versindia.org. Case study.

Naming: Skopaq is the company. Argus is the product. Some repository names still carry older branding; "Skopaq" is never a product name.

Only repositories with a licence file that GitHub detects are listed. Code published without one is not open source, whatever its README says, so it is not called that here.

Not every public repository here is a product. Several are reference projects, forks or earlier work that the company no longer publishes on skopaq.com. They remain public as they are; nothing about them has been withdrawn.

Built by agents. Gated by machines. Approved by people. Six layers, each saying what the agents do, what the machine refuses, and where a person signs. The full version is at skopaq.com/method.

1. Intent — agents draft the specification, plan and blast radius. Nothing is built from an inferred requirement; a person approves the plan before any agent starts.
2. Architecture — schema-first contracts, generated API surfaces and infrastructure. Validation answers with named errors. Cloud, edge or fully air-gapped, per product.
3. Build — specialised agents in sandboxed worktrees over a typed tool bus, enforced statically. No direct writes to main. A person is deliberately absent from this loop; review happens at the gate.
4. Verification — every diff meets the same gates: documentation, types, lint, format, tests, build. Any failure blocks. A release with an ungrounded claim fails the citation check; seeded replay and golden outputs; a person owns the reference sets the checks compare against. These are our own checks, not a regulatory compliance service.
5. Operation — systems watch health, budget and drift, patch what they can, trip a breaker on a runaway loop, and escalate to a person for decisions. Not an availability commitment: no SLA, no on-call rota, no 24/7 support offering.
6. The answer — every claim carries its source or a refusal is served. Advisory only; synthetic reference data for rehearsal; deterministic replay. A person executes every step.

In practice, four things a reader can verify rather than take as values:

• The spec comes first. Intent, schemas and boundaries are written and approved before implementation starts. An agent asked to infer the requirement will invent one.
• Agents do the work; gates decide if it ships. What reaches main is decided by tests, static guards and a deterministic merge pipeline, not by whether the diff looked reasonable.
• A claim that cannot be checked does not ship. This applies to the products and to the website. Where something cannot be substantiated it is absent, and there is usually a comment in the source saying why.
• Small surface, deliberately. Hand-rolled where a library would have been quicker; we would rather own the code than inherit its maintenance.

From skopaq.com/trust, which exists so a security or procurement review can read it in one place.

• No SOC 2 attestation, no ISO 27001 certification, and no claim of HIPAA certification or compliance. Where an engagement needs a security review, we answer it directly.
• No benchmark score. VELA targets the gap the ESA-ADB benchmark identified and has not yet been run against it.
• No client name, logo, quote or testimonial without that client's written consent. One client engagement is published, with consent; there are no others.
• No figure — an accuracy, a saving, a percentage — that we cannot show you the source of.
• No availability commitment: no published SLA, on-call rota or 24/7 support offering.

We write down what we remove. The site keeps a public account of the claims deleted from our own marketing and why — What we removed from our own website, and why — alongside Why provenance, not confidence.

• General, security, press and careers: solutions@skopaq.com. A person answers.
• Billing, invoices, GST and refunds: finance@skopaq.com.
• Security reports: email with [SECURITY] at the start of the subject line, or use a repository's Report a vulnerability button where it is enabled. The policy is in SECURITY.md and, in RFC 9116 form, at skopaq.com/.well-known/security.txt.
• Careers: there is no open hiring process today, and we would rather say so than post roles that do not exist. Write anyway if this is the work you want to be doing — skopaq.com/careers.
• Product enquiries: skopaq.com/contact.
• The corporate record — CIN, GSTIN, registered office, jurisdiction — is at skopaq.com/company#corporate. The press kit is at skopaq.com/press.