AI Research

An AI innovation lab that also trains engineers.

Our students don't just learn AI—they contribute to our active research projects. Every course track feeds into a live system with real users, real latency budgets and real failure modes.

Explore courses Talk to the research team

India-first by design

We build for Indian languages, accents, infrastructure and price points — not as a localisation afterthought.

Grounded and auditable

Every agent output is traceable to a tool call or source document, with logging, guardrails and consent baked in.

Shipped, not shelved

Research is judged by whether it survives production traffic, latency budgets and real users.

Students as contributors

Learners join active project pods, review each other's code and ship components used in the live systems.

Active projects

Research in progress

Two flagship systems in build today, both staffed by MindForge engineers alongside student pods.

Indian Language Emotional AI Calling Agent

Research in Progress

A multilingual voice agent for Indian enterprise contact centers with emotion detection and voice cloning.

Indian contact centers handle millions of calls a day across a dozen languages, and off-the-shelf voice AI collapses the moment a caller code-switches between Hindi and English mid-sentence. We are building a speech-to-speech agent trained on Indian accents and mixed-language speech that listens, detects the caller's emotional state in real time, and adapts its tone, pacing and escalation behaviour accordingly — so an irate customer is routed to a human before the conversation breaks down.

HindiMarathiGujaratiKannadaTamilEmotion DetectionVoice CloningHuman-like Conversation

Multilingual ASR for Indian speech

Fine-tuning open speech models on accented, code-switched Indian audio so recognition holds up on noisy mobile calls.

Real-time emotion & intent detection

Prosody and text signals fused to classify frustration, confusion and urgency within the first few seconds of speech.

Voice cloning with consent controls

Brand-consistent Indian-language voices generated from short consented samples, with watermarking and usage logging.

Barge-in and latency engineering

Sub-second turn-taking, interruption handling and graceful fallback so conversations feel human rather than scripted.

Stack

Whisper-class ASRLLM dialogue plannerNeural TTSWebRTC / SIP telephonyVector memory

Where students contribute

Students annotate and clean regional speech datasets, run evaluation harnesses on emotion classifiers, and build the call-analytics dashboards used to review agent transcripts.

Agentic Travel Planner

Demo Coming Soon

Multi-agent system that plans end-to-end trips—flights, hotels, visas, weather and personalised itineraries.

Travel planning is the ideal stress test for agentic AI: it requires live data, hard constraints like budget and visa eligibility, and dozens of interdependent decisions. Our planner decomposes a single natural-language request into specialised agents that search, negotiate trade-offs, and reconcile conflicts through a supervising reasoning loop — then returns a day-by-day itinerary with citations for every booking option it proposes.

Flight SearchHotel BookingVisa AssistantBudget PlanningWeatherMulti-Agent Reasoning

Supervisor–worker agent topology

A planner agent decomposes goals and delegates to flight, stay, visa and weather workers, then merges their results.

Tool use and grounded retrieval

Every recommendation is backed by a live API call or retrieved document — no invented prices or fabricated rules.

Constraint solving under budget

Cost, travel time and comfort traded off explicitly so the itinerary respects a stated budget ceiling.

Evaluation of agent trajectories

Automated scoring of plan quality, tool-call efficiency and failure recovery across a benchmark of trip briefs.

Stack

LangGraph-style orchestrationFunction callingRAG over visa & policy docsCaching layerTrace evaluation

Where students contribute

Students own individual worker agents, write the tool wrappers and guardrails, and contribute test cases to the trajectory evaluation suite.

Method

How a project moves from idea to production

Problem framing

A real Indian industry pain point, scoped with a measurable success metric.

Data & baselines

Dataset collection, annotation and a dumb-but-honest baseline to beat.

Prototype pods

Small student-plus-engineer pods build competing approaches in parallel.

Evaluation harness

Automated benchmarks for quality, latency, cost and failure recovery.

Production hardening

Guardrails, observability, cost controls and human-in-the-loop fallbacks.

Publish & open source

Findings written up, components released, students credited.

Roadmap

Upcoming verticals

Domain-specific agent systems we are scoping next, each with an industry partner and a student pod.

Medical AI

Clinical documentation, triage assistants and diagnostic decision support.

Legal AI

Contract review, case-law retrieval and drafting copilots for Indian law.

Education AI

Adaptive tutoring, assessment generation and learner analytics.

Factory AI

Vision-based defect detection, predictive maintenance and shop-floor agents.

Retail AI

Demand forecasting, catalogue enrichment and conversational commerce.

Government AI

Citizen-service agents, multilingual grievance handling and document processing.

Join a research pod

Learn AI by building it with us.

Research pods are open to learners across our tracks. Tell us what you want to work on and we will map you to a project.