The system
Aminobots builds the same three products for any industry, running on one engine:
- Assist — act. A multi-channel operations agent that drafts, acts, and escalates across WhatsApp, voice, web, email, and Slack. A human takes the edge cases.
- Lens — decide. Reads your operational signals, reasons over them, and takes the next step. It deploys as the named application your industry understands: Helm (lending), PowerIQ (industrial energy), KidneyCare (clinical), MIS (reporting).
- Nous — compound. A knowledge engine that searches and reasons across your corpus and answers with citations, powering L1/L2/L3 support and getting more useful as your knowledge grows.
- Blueprint — the engine. The agent compiler underneath every deployment. One spec targets AWS, Azure, or GCP with LLM routing, safety gates, and an audit trail baked in.
How deployment works
Every system runs inside the customer's own cloud tenancy — the customer's identity, keys, and data residency. Nothing leaves the boundary the customer already audits; an on-prem option is available and the architecture is DPDP-ready.
Engagements follow a fixed shape: Diagnose (a 2–3 week written scorecard — is AI the right move, where's the return, what to build first) → Deploy (build, ship, handover) → Operate (observability, upgrades, audits) → License (your team runs it). Before any code, the build passes a Gate 0–4 discipline: do you even need AI, keep it simple, shape it to the workflow, measure it in your control, and build it to connect.
Live case studies
Four deployments are live in production. Each is summarised below with the problem, the approach, the stack, and the measured results.
Helm at IFHL — WhatsApp loan intake to disbursement
Problem. A loan that took days to move from application to disbursement, with manual checks and hand-offs — while faster competitors won the customer.
Approach. An agentic flow (deliberately a flow, not a peer mesh) runs the whole journey over WhatsApp: intake, parallel verification across systems, a one-screen decision, and disbursement — with a human only on the edge cases.
Architecture & stack. Lens deployed as Helm on the Blueprint engine, inside IFHL's own cloud; ~22 APIs orchestrated (KYC, bureau, banking, disbursement); WhatsApp as the customer channel; full audit trail.
| Metric | Result |
|---|---|
| Application → disbursement | ~8 minutes (WhatsApp to money) |
| Turnaround time | ~50% faster |
| Straight-through processing | 99%+ |
| Cost per case | ~30% lower |
| APIs orchestrated | 22 |
TRACE at GyanDhan — support across 2,000+ WhatsApp groups
Problem. Customer questions spread across 2,000+ WhatsApp groups, with response times that stretched to hours and a support team that couldn't scale with them.
Approach. An agent answers questions autonomously across every group, using a knowledge core with citations, and hands the genuinely tricky ones to a human.
Architecture & stack. Assist for the channels plus Nous for the knowledge, on the Blueprint engine, in the customer's cloud.
| Metric | Result |
|---|---|
| Automation rate | 99%+ handled without a human |
| Cost per query | ~$0.02 |
| Response time | ~2.5 hours → ~15 minutes |
| Coverage | 2,000+ WhatsApp groups |
PowerIQ at Omega Transmission — cutting the HT electricity bill
Problem. A plant paying a flat rate regardless of when it drew power — an electricity bill that didn't care about time-of-day tariffs — with heavy machines running at the most expensive hours.
Approach. Five-minute metering feeds a decision agent that shifts heavy machines to cheaper time-of-day windows automatically, and produces ESG Scope-2 reporting on demand. Meter data is processed at the edge so it never has to leave the site.
Architecture & stack. Lens deployed as PowerIQ with an edge component (Raspberry Pi on site) rather than moving raw meter data to the cloud.
| Metric | Result |
|---|---|
| HT electricity bill | ~20% lower |
| Metering granularity | 5-minute intervals |
| Time-of-day optimization | 100% |
| ESG Scope-2 report | ~60 seconds |
KidneyCare at AIIMS Patna — a phone that reads a dipstick
Problem. Early kidney-disease screening needs a pathology lab and a trained eye — neither of which is at the point of care, especially for children in under-served settings. The test strip costs about ₹12; the constraint is reading it reliably.
Approach. A phone camera reads a urine dipstick in seconds; a computer-vision model grades the pads and flags early kidney-risk. A 70% confidence floor means the system says "indeterminate" rather than guess on a child.
Architecture & stack. Lens deployed as KidneyCare with on-device inference; built and validated with the AIIMS Patna paediatrics team, progressing along the CDSCO regulatory pathway.
| Metric | Result |
|---|---|
| Accuracy vs pathology lab | 90%+ (clinical validation, n=30) |
| Cost per screening | 70–80% lower |
| Photo → graded result | ~11 seconds |
| Maturity | TRL 4→6 · CDSCO pathway |