Under The Hood

What's under the hood?

The services, engines, and open-source projects that Ananta runs on — with a plain-English note on what each one does for you.

Ananta is built and operated on around 14 interconnected services. Some are open-source projects we lean on; some are engines we wrote ourselves. Every response is grounded in real web searches, every action is logged, and every engine here is either open-source or something we built and can inspect line-by-line.

The conversational brain

DeepAgents is LangChain's open-source Python agent harness (langchain-ai/deepagents) — the layer that composes responses and decides which tools to call. We run it on Grok-4-Fast-Reasoning via AIMLAPI.

Sarvam 105B is a multilingual LLM built for Indian languages. When the conversation is in Telugu, Hindi, Tamil, Bengali and so on, we route through Sarvam for a native touch that generic models simply don't have.

asymm-intelligence is a pure-Go routing and memory service we wrote in-house. It handles intent classification, emotional tone detection, and conversation memory before a single token is spent on an LLM — which keeps fast interactions cheap and fast.

Your documents

Stirling PDF is the open-source PDF workhorse. Merging, splitting, compression, page rotation, extraction, and OCR across English, Hindi, Telugu, and Arabic.

Gotenberg with LibreOffice under it handles Word, Excel, and PowerPoint to PDF conversion — Docker-friendly, battle-tested, boring in the best way.

MathAlive is our in-house engine that turns plain natural-language instructions into polished PDFs. When you say "invoice for 5 licenses at 1000 BHD with 18% GST", it's MathAlive on the other end assembling the document.

Web grounding

SearXNG is a privacy-focused metasearch engine. It aggregates results from multiple sources without passing your query to trackers. This is how we ensure every grounded answer has real sources attached — no ads, no re-selling of your intent.

Crawl4AI is our headless-browser web crawler. When you paste a URL and ask for a summary, this reads the page and extracts the meaningful content.

Voice and vernacular

Sarvam Saaras:v3 is the speech-to-text model we use, built for Indic languages. It transcribes voice notes natively in the language you spoke — no English detour. 23 languages supported.

Sarvam Bulbul:v3 is the matching text-to-speech model when Ananta speaks back. 35+ voices across 11 languages.

Data, storage, and delivery

SpacetimeDB is the real-time synchronised database we use for reliable message delivery. If your network blinks, your messages queue up and sync when it recovers — they don't get lost.

MinIO is S3-compatible object storage running on our own hardware. When you upload a file, it lives on our infrastructure — not Amazon's, not Google's — and we keep it only as long as you need it.

Postgres 16 (we call our self-hosted instance asymm-db) is our operational database for accounts, credits, and state.

Caddy sits at the edge. TLS termination (HTTPS), HTTP/3, graceful routing between services, automatic certificate renewal.

Our own engines

Asymm Engine is a .NET 10 service with 70+ mathematical kernels. It houses our OCR intelligence (PredatorVision, VedicValidator, SparseOCR+DNA), our ERP logic (invoice, payment, ledger, Tally bridge), and a quaternion-S³ geometric search algorithm you won't find anywhere else. This is the part that's purely ours.

MathAlive (mentioned above) and asymm-intelligence (also above) round out the in-house trio. Everything else is open-source software we respect, configure carefully, and contribute back to when we can.

Infrastructure principles

  • Every engine here is either open-source or something we built ourselves. No closed black boxes in the middle of your data flow.
  • File storage is on our servers — not a public cloud owned by someone else.
  • Everything is health-checked and logged. If a service misbehaves, we see it before you do.
  • Graceful degradation — if one piece fails, the others keep working.
  • When we update, we update. No silent redeployments that change behaviour without you noticing.

Acknowledgments

We are deeply grateful to the open-source projects that make Ananta possible. A partial list:

If one of these has been especially useful to us, we try to contribute back — bug reports, patches, or sponsorship where we can. If you maintain one of these and we haven't reached out, please do say hello.

Last updated: 2026-04-18