Intelligent Automation
The React → Act → Heal loop, automated by specialist models — not one generalist trying to do everything. Each Avirka model has one job and is trained for it end-to-end.
Coming 2026 Q3 · Apache 2.0 · Self-hostable
SRE LLMs that hypothesize, act, and heal.
Avirka™ is an open-source LLM family by WarbleCloud, purpose-built for the React → Act → Heal incident-response loop — not another general-purpose model retrofitted for ops.
Small enough to run on a single H100. Tuned on real postmortems
and chaos-engineered failure data. Drops into your stack via the
LLM_PROVIDER=avirka flag.
Capabilities
Three things Avirka does better than a frontier model.
Predictive Intelligence
Trained on real postmortems, runbooks, traces, and synthetic chaos failures. Avirka hypothesises root causes the way your senior SRE does — because it learned from how they actually write incident reviews.
Seamless Integration
One env var (LLM_PROVIDER=avirka) to swap from Vertex, Groq, or Anthropic. Stay mixed per workload — frontier models for prose, Avirka for ops. No vendor lock-in either direction.
In action
Built to evolve. Designed to ship.
Experience an SRE assistant that hypothesises causes, gates risky actions, and writes the post-mortem — without leaving your cluster. Designed to evolve as your incident catalogue grows.
Demo video — coming with v1 (2026 Q3)
Works with
Kubernetes
Prometheus
Grafana
OpenTelemetry
vLLM
HuggingFace
ChirpStack
Warble Cloud
PagerDuty
Slack
The lineup
One family, five specialists.
Each Avirka model has one job in the loop. Compose them, or pick the one you need.
Actor
avirka-react
Generates incident hypotheses from metrics, logs, and runbooks. The default Actor in the Reflexion engine.
Llama 3.1 8B + LoRA
Critic
avirka-critic
Safety gate. Scores hypothesis confidence and blast radius before any action lands. Trained on hypothesis-outcome pairs.
Llama 3.1 8B + LoRA
Heal
avirka-heal
Post-mortem RAG. Retrieves runbooks and prior fixes, drafts remediations, and writes the incident review.
Qwen 3.5 (compact)
FinOps
avirka-finops
Cost-anomaly explanation. Reads BigQuery + AWS billing diffs and turns "$2k spike" into a one-paragraph why.
Llama 3.1 8B + LoRA
Chat
avirka-chat
Cockpit assistant. Answers SRE questions over your live cluster context and tool-calls when needed.
Llama 3.1 8B + LoRA
2027 Q1
avirka-mix
Phase-2 MoE replacement that subsumes the family. Qwen 3.5 base, 17B active / 397B total. Production parity with Vertex.
Qwen 3.5 MoE
Why Avirka
Built for the only workload that matters at 3am.
01
Specialist beats generalist
GPT-4-class models are trained to write essays. Avirka is trained on postmortems, runbooks, alerts, traces, and synthetic chaos data. On the React/Act/Heal loop it matches Gemini 2.5 Flash at a fraction of the cost.
02
Self-hosted by design
Apache 2.0 weights. vLLM serving on a single H100, or a fleet of L4s. Your incident data — and the embeddings it produces — never leave your cluster. Air-gapped deploys supported via WarbleCloud module.
03
Per-tenant LoRA
Your runbooks. Your service catalog. Your past incidents. Avirka's per-tenant LoRA pipeline fine-tunes a private overlay on top of the public base — without ever shipping your data outside.
04
Drop-in via LLM_PROVIDER
If you already run Reflexion, switching is one env var:
LLM_PROVIDER=avirka. Stay on Vertex, Groq, or Anthropic for
workloads that benefit from a frontier model. Mix and match per phase.
05
Eval-driven, not vibes-driven
Every Avirka release ships with a regression report against frontier baselines: hypothesis acceptance, time-to-resolution on replayed incidents, cost-per-incident, false-positive rate in shadow mode.
06
Open weights, open recipes
We publish the LoRA recipes, the synthetic data generators, and the eval harness. Fork it, tune it, run it on your hardware. Phronix exists for teams who want a contract instead.
Roadmap
Four phases. Twelve months.
P0 · 2026 Q3
avirka-react v1
- Llama 3.1 8B base + LoRA on public postmortems
- Synthetic incidents from Litmus + Chaos Mesh
- vLLM serving on GKE
- Shadow-mode rollout in Warble Cloud
P1 · 2026 Q4
avirka-critic + per-tenant LoRA
- Critic LoRA from hypothesis-outcome feedback
- Per-tenant fine-tune pipeline (private data)
- Cockpit toggle: shadow → live
P2 · 2027 Q1
avirka-mix (Qwen 3.5 MoE)
- Multi-node H100 fine-tune
- Production parity with Vertex DSQ
- 10× cost reduction validated in production
P3 · 2027 Q2
Open-weights release
- Apache 2.0 on HuggingFace under
warblecloud/org - Reference fine-tune recipes published
- WarbleCloud module for one-command self-host
- Phronix tier announces (sister commercial product)
FAQ
Answers, before you ask.
How does the Avirka waitlist work?
Tell us your name, email, GitHub or X handle, and your top SRE pain in one line. We email exactly once — the day v1 weights land on HuggingFace. No drip campaign, no marketing follow-ups.
What is Avirka?
Avirka is an open-source LLM family built by WarbleCloud LLC (USA) and Chirpstack LLP (India), purpose-trained for site reliability engineering. Avirka™ is a registered trademark of both entities. Five specialist models (react, critic, heal, finops, chat) cover the React → Act → Heal incident-response loop. Small enough to run on a single H100.
Is there a cost to join the waitlist or run Avirka?
The waitlist is free. Avirka model weights will be released under Apache 2.0 — free to download, self-host, and modify (the Avirka™ name itself remains a registered trademark of WarbleCloud LLC and Chirpstack LLP). You pay only for the GPUs you run them on. Warble Cloud (the hosted version) is the commercial path for teams that don't want to operate the infrastructure.
What platforms does Avirka support?
Anywhere vLLM runs: bare-metal Linux, Kubernetes (GKE, EKS, AKS, on-prem), or air-gapped data centres. NVIDIA H100 is the reference GPU; L4 fleets also supported for the smaller models. Integrates with Prometheus, Grafana, OpenTelemetry, and Kubernetes-native incident tooling.
Will I get early access if I join the waitlist?
Yes — the launch email contains the HuggingFace links the day weights go public. Waitlist members also get first access to the per-tenant LoRA fine-tuning pipeline, which lands in 2026 Q4.
Who is Avirka for?
SREs, platform engineers, and on-call teams who want LLM-assisted incident response without sending production telemetry to OpenAI, Anthropic, or Google. Particularly suited to regulated industries (finance, healthcare, defence) where data residency rules out frontier APIs.
How does Avirka compare to GPT-4 or Gemini for incident response?
Frontier general-purpose models are trained to write essays; Avirka is trained on postmortems, runbooks, alerts, traces, and chaos-engineered failure data. On the React/Act/Heal loop Avirka matches Gemini 2.5 Flash at roughly 10× lower inference cost. Frontier models still win on long-form prose — we recommend mixing per workload.
When does Avirka launch?
Phase 0 (avirka-react v1) ships 2026 Q3. Phase 1 (avirka-critic + per-tenant LoRA) lands 2026 Q4. Phase 2 (avirka-mix MoE on Qwen 3.5) ships 2027 Q1. Open-weights release under Apache 2.0 is 2027 Q2.
How is Avirka different from Phronix and Warble Cloud?
Three layers from the same group: Avirka™ is the open-source LLM core (trademark-registered to WarbleCloud LLC in the USA and Chirpstack LLP in India), Phronix is the proprietary reasoning and decision subsystem that sits on top of Avirka, and Warble Cloud is the hosted SaaS that ships Avirka + Phronix as a managed product. You can self-host Avirka standalone for free, or use the full stack as a service via Warble Cloud.
Will Avirka model weights actually be open-source?
Yes — Apache 2.0 weights are the public commitment for the 2027 Q2 release. Pre-release artifacts (current weights, training data, recipes) remain proprietary to WarbleCloud LLC and Chirpstack LLP until then. The Avirka™ name itself remains a registered trademark of both entities even after the weights ship under Apache 2.0; use of the name in derivative work is governed by the Avirka Trademark Policy.
The name
What "Avirka" means.
Avirka (pronounced a-VEER-ka) is a coined name — fully invented for the brand, not borrowed from an existing word. That intentional coinage is what gives the trademark its defensibility under USPTO and Indian TM law: a name with no prior meaning has the strongest exclusive claim.
A founders' blend
The name is a private blend of the two founders' names: AV from Avinash, and IRKA (the letters I, R, K, A) from Rasika. We wanted the technology we build together to carry our names in a way that wasn't obvious to outsiders — a small private joke that sits underneath the public brand.
A happy phonetic echo
Read out loud, a-VEER-ka rhymes with the Sanskrit root vīra (वीर) — meaning brave, strong, heroic. Adding the soft -ka diminutive that Sanskrit uses for the small or the personal gives a reading like "the brave little one."
That meaning wasn't the source of the name — but it turns out to fit what Avirka is built to do. A small specialist model that shows up at 3am when the page goes off, hypothesises the cause, gates the risky action, drafts the fix, writes the post-mortem. Not a frontier giant. A focused, brave little one.
A pair with Frakma
Frakma, the working product-family codename, phonetically pairs with Avirka — both coined, both three syllables, both ending in an open a sound. The family rhymes; the products belong together.
The trademark
Avirka™ is a registered trademark of WarbleCloud LLC in the United States and of Chirpstack LLP in India. Use of the name is governed by the Avirka Trademark Policy: factual reference is fine ("my stack runs Avirka", "Avirka-compatible"); using "Avirka" in the name of a derivative model or competing product is not. We chose a fully coined name precisely so that derivative-naming rules are clear and enforceable for everyone.
Get notified the day v1 ships.
No drip campaign, no marketing emails — one note when the first Avirka weights land on HuggingFace.
Protected by Cloudflare Turnstile. Your details flow to WarbleCloud's signup database only — never sold, never used for ads.