AI capabilities
7 CPU models on day one. 6 GPU models from Phase 3 Month 11. No tier gating, no per-call limits, no third-party AI in the critical path.
Strategy
The AI stack is opinionated on purpose. These are the opinions.
01
Eight CPU models — scoring, forecasting, NER, dedupe, sentiment — work on day one on commodity infrastructure. The RTX 4070 Ti Super 16GB lights up in Phase 3 Month 11, once we have the client volume to amortise it.
02
The AI Reminder Engine takes the buyer-risk score and turns it into a prioritised Action Queue with KPI math behind it. AI does not chat at your sales team — it tells them who, when, and what.
03
Qwen3, BGE-M3, RoBERTa, spaCy, Faster-Whisper — all open weights, all running on our infrastructure inside India. No data leaves the country. Groq API stays as a fallback for early-load smoothing.
The thirteen
Each tile lists the model, the compute target, the latency budget, and the job it does inside the product.
XGBOOST · CPU · ~1MS
A 0–100 score on every lead, retrained weekly. Sales sees the hot list first, not the new list.
XGBOOST + PROPHET · CPU · MINUTES
Project-wise monthly bookings forecast with seasonality. The MD's dashboard answers next quarter.
XGBOOST · CPU · MINUTES
Per-demand probability of on-time payment. Finance knows what to actually expect this month.
XGBOOST · CPU · ~5MS
Risk score per active SO. The AI Reminder Engine prioritises the buyers most likely to slip.
RAPIDFUZZ · CPU · <10MS
Fuzzy phone + email match at lead ingest. One buyer, one record — even when the call centre misspells.
SPACY · CPU · ~20MS
Pulls names, projects, unit types, budgets from inbound email and WhatsApp before they hit a human.
ROBERTA · CPU · ~50MS
Sentiment on every transcribed call and WhatsApp thread. Negative trends become tasks automatically.
BGE-M3 · CPU/GPU · ~80MS
pgvector-backed similarity search. "Find me units like this" works the way buyers actually ask.
FASTER-WHISPER · GPU · REAL-TIME
Every CTI recording becomes searchable Telugu/Hindi/English text within seconds of the call ending.
QWEN3 4B · GPU · ~2S
Three bullets per call: what the buyer wants, what was promised, what the next step is.
QWEN3 8B · GPU · ~15S
First-cut agreement generated from the SO using your lawyer's template. Review, redline, e-sign.
QWEN3 4B (THINKING) · GPU · ~5S
Eligibility estimate from income + CIBIL + obligations. Buyers learn the answer in the first conversation.
QWEN3 4B + LLAMAINDEX · GPU · ~3S
ProXA 24×7 buyer assistant grounded on your agreements, brochures, and FAQs. ProX only — Customer Portal required.
GPU server
₹1,40,000 one-time at Month 11, when client count crosses fifteen. 16GB GDDR6X holds the full Qwen3 + BGE-M3 + Whisper stack with about 7.5GB of headroom. Cloud GPU costs more after three months, and pricing volatility is real. Owning the box wins.
Data residency
Schema-per-tenant Postgres. Realm-per-tenant Keycloak. Per-tenant MinIO prefix. Models run on our hardware, in our colocation. No third-party AI processor in the call path, no buyer PII shipped to a foreign vendor. The audit-log triggers prove it on demand.
Ready when you are
A live walk-through tailored to your project pipeline, your payment plans, and your team. Forty-five minutes.