Where our AI is trained, benchmarked, and invented — in the open. Live numbers below come straight from production.
Register-preserving replies — 100% on our benchmark via measured prompt engineering. Try: chat.indiacrm.in
Execution-tested: working HTML pages, forms, SQL, security-pattern fixes. Try: coder.indiacrm.in
Any audio format, OpenAI-compatible endpoint, served from Bangalore.
Hinglish description → designer-grade landing page wired into a CRM lead pipeline, on your own domain.
Grounding took GST accuracy 16.7% → 100% — self-curating knowledge cards keep it current.
Bol ke pucho: hold-to-speak Hinglish voice AI in any phone browser — our whisper STT (keyless) + smart cascade + on-device TTS. The mass-adoption form factor, free. Try: voice.indiacrm.in
Full VS Code in the browser at ide.indiacrm.in — Open VSX plugin ecosystem plus our own IndiaCRM AI extension: chat, refactor, and ghost-text completions on our Bangalore engines.
Design-first app builder at coder.indiacrm.in: mockup for approval before any code, file manager, live preview, chat-to-iterate.
Our 7B-class engine lit up 18 July after a server upgrade — 85.2% on IndiaCRM-Bench, Hinglish register 100%. Try: max.indiacrm.in / secure.indiacrm.in
A research harness runs continuous experiments on our servers — decoding parameters, engine cascades, resampling policies. Every trial is logged, failures included. Findings are confirmed only after beating baseline by ≥5 points and replicating.
Loading trials…
Our quality gates mechanically manufacture DPO preference pairs (register-defecting reply = rejected, passing retry = chosen). Zero human labelling. Training signal accumulates nightly.
The model proposes improvements to its own reference cards; a benchmark judges; only measured wins are promoted. The knowledge base evolves like code — versioned and audited.
Round-the-clock champion/challenger loops over prompts, cards and data — ~100 measured experiments a day, yielding only when real users need the CPUs.
Coming: training on what actually worked — replies that got answered, quotes that closed, COD calls that prevented returns. Only the owner of the system of record can do this.
Our published benchmark, including two first-of-their-kind sections: messy Indian address parsing and GST business knowledge. Every engine is judged by it, win or lose.
Our inventions stacked: bench-informed routing (each task to the measured-best component), the self-trained Nano for addresses, card-grounded 7B for GST (100%), few-shot register anchoring (100%), validator net. Beats every raw engine we run — the stack IS the model.
Our routing invention: the fast engine answers, free mechanical validators judge the reply after generation, only failures escalate to the 7B. Small-model speed, big-model quality floor — two positive runs (+6.4, +1.4 — magnitude under study). Ships as indiacrm-auto.
First checkpoint trained by IndiaCRM, in India, on our servers and curated data — fully automatic pipeline. Scored 59.6% vs the 70% incumbent: NOT promoted (the gate worked), but address parsing jumped +14 and entities +4 where our data is dense. The regression diagnosed a training bug, already fixed for the next automatic run. Losses published, always.