Described at a level our clients are comfortable with.
Most of our AdTech engagements sit under client confidentiality, so the studies below describe the problem, approach, and outcome without naming the account. Where we can point to public evidence — like our plugin catalog — we do.
Programmatic infrastructure
Header bidding & SSP/DSP integration for a Revive Adserver-based programmatic platform
Problem: International DSP/SSP operators running Revive Adserver needed OpenRTB-compliant supply- and demand-side integrations, header-bidding, and marketplace transaction logic — the operational backbone of a high-concurrency ad exchange.
System: Custom-built integrations and transaction logic layered directly onto their existing Revive Adserver deployment.
Automation: Bid request/response handling, header-bidding coordination, and marketplace settlement logic running without manual intervention per transaction.
Outcome: A high-concurrency, real-time system where coordination failure has an immediate revenue cost — the operating experience the AMEI framework was later formalized from.
AdCP integration for an AI-native advertising platform
Problem: An AI-native advertising platform needed to automate campaign discovery, negotiation, and activation across CTV and digital placements, without relying on a traditional programmatic pipeline.
System: An AdContext Protocol integration connecting natural-language advertiser intent directly to publisher inventory.
Automation: Agent-to-agent negotiation between advertiser and publisher agents, logged and auditable at every step.
Outcome: Agentic commerce in advertising, working — replacing a manual media-buying workflow with direct machine-to-machine negotiation.
Problem: Campaign managers on an established ad-serving platform had to manually cross-reference live metrics against guideline documents to answer routine questions.
System: A conversational front door layered onto the platform's existing data infrastructure, with no migration required.
Automation: Plain-English questions answered by blending live metrics with guideline documents, with anomalies flagged automatically.
Outcome: Delivered in four weeks, with every answer sourced back to its underlying data — no black-box responses.
Real-time invalid-traffic detection on the bid path
Problem: Real-time bidding traffic needed fraud and invalid-traffic scoring without adding latency to the live serving path.
System: A production IVT plugin combining rule-based weighted scoring with an unsupervised anomaly-detection layer.
Automation: Every incoming request scored automatically and routed to a Clean / Suspicious / Blocked verdict before serving continues.
Outcome: Verdicts fast enough to sit directly on the bid path without adding noticeable latency — no manual review queue for the vast majority of traffic.
We've also developed an early-stage agritech/foodtech concept for an AI-enabled smart cooking platform — pre-portioned ingredient kits paired with app or QR-based cooking profiles, sensor-driven monitoring, and AI-adjusted timing and temperature to standardize a process that's traditionally dependent on manual judgment. Still at concept and prototype-design stage, it's a useful proof that our orchestration-and-verification thinking isn't specific to advertising.
Five AI-native plugins from this practice (demo available under NDA) are live at reviveadservermod.com — alongside our wider catalog of 300+ AdTech plugins, with 1,000+ plugins sold to 500+ clients across 180+ countries, the operating base this AI work builds on.
This work is grounded in our applied research paper — Autonomous Programmatic Clearing — which benchmarks the verification pipeline these production plugins implement.
Selected builds
Systems built by our engineering team and delivery network.
The same patterns, governed tools, typed contracts, validation loops and human approval, run across these builds in finance, healthcare, travel, HR, operations and industry. Client names and figures stay confidential.
Multi-agent orchestration and MCP
Build
What it does
Stack
MCP agentic operations platform
A multi-agent platform that discovers tools dynamically, uses typed contracts and routes consequential steps through human approval. Enterprise databases, APIs and documents are exposed as governed tools.
LangGraph, CrewAI, MCP, Pydantic
Self-correcting enterprise agent pipelines
Agent pipelines on custom MCP servers with automatic retry and validation loops, then escalation to a person, plus structured enterprise data integration.
Custom MCP servers
Seven-agent software development studio
Coordinated agents for context analysis, planning, validation and execution, working across several model providers.
Anthropic, OpenRouter, Ollama, Gemini
Multilingual agentic financial workflows
Production workflows with retrieval and MCP tool calling over enterprise SQL databases, in use by a large active user base.
LangGraph, RAG, MCP, async FastAPI
Agentic assistant with persistent memory
A multi-tool assistant using MCP over stdio and HTTP, with memory across sessions and real-time streaming.
LangGraph, MCP, SQLite
Multi-agent planning and delegation
Breaks complex work into tasks, delegates them and verifies structured outputs before they are accepted.
Validation loops, structured outputs
Knowledge systems and domain copilots
Build
What it does
Stack
Voice receptionist agent
A stateful voice agent that schedules appointments and answers from approved knowledge, tracking the conversation as it goes.
LangGraph, OpenAI, Postgres, pgvector, FastAPI
Financial document search
Hybrid dense-vector and keyword retrieval over large sets of complex financial documents, with high relevance on RAGAS evaluation.
Vector search, BM25, FastAPI, Streamlit, RAGAS
Financial copilot and operations assistant
Multi-agent financial analysis and transaction routing, plus an assistant that answers from internal SOPs.
LangGraph, PostgreSQL, Groq, pgvector
Healthcare data assistant
Plain-language data access through a custom MCP server. Calculations run deterministically in the database, not in the model.
MCP, SQLite, multiple model providers
Domain agent workflows
Multi-agent workflows for travel, job search and finance, with caching, persistent chat history and streaming responses.
Google ADK, Gemini, Redis, SSE
Invoice and business document AI
Layout-aware parsing and invoice data extraction, with plain-language business queries.
Qwen, LangChain, FAISS, Document AI
ERP/CRM assistant on self-hosted models
Large-scale document ingestion and query resolution with open-source models on dedicated GPU servers.
Llama 3, DeepSeek
SOP and onboarding assistant
One search point for internal procedures and onboarding documents.
FastAPI, pgvector
Evaluation, monitoring and domain workflows
Build
What it does
Stack
Agent benchmark framework
Containerised environments with automated grading, to evaluate AI coding agents reproducibly.
Docker, automated grading
Predictive fault detection with root-cause summaries
Anomaly models flag pipeline faults, and a local model drafts the root-cause summary for engineers, with dashboards.
Isolation Forest, Random Forest, Qwen via Ollama
Personalised email automation
Agents extract data from CSV and PDF files and fill personalised templates.
CrewAI
Travel and medical Q&A
Multi-turn planning and question answering within domain-specific guidelines.
LangChain, LangGraph, Gemini
Local scientific document search
Semantic search and Q&A with local embeddings and local models.
Ollama, LangChain
Text classification and document Q&A
A deep-learning classifier combined with vector retrieval.
BiLSTM, vector search
Computer vision and speech
Build
What it does
Stack
Speech intelligence and speaker diarization
Real-time transcription that identifies who spoke when.
Whisper, SpeechBrain, Resemblyzer
Multilingual transcription
Audio and video transcription across multiple regional languages.
faster-whisper, Gradio
Visual defect and anomaly detection
Automated visual inspection and quality screening for industrial and operational workflows.
TensorFlow, ResNet50, MobileNetV2
Image classification pipeline
Image preprocessing, feature extraction and multi-class classification.
As an AI-first lab, we occasionally take on product concepts outside AdTech that fit the same pattern — structured experimentation, sensor or data feedback, and AI-driven decision logic. One current example, still at an early concept stage, is an IoT-connected smart cooking platform that uses AI-driven profiles to standardize a manual, judgment-heavy process, developed through our COXBIT innovation-lab connection at TNAU Coimbatore. It's a small part of what we do, but a useful proof that the underlying architecture isn't specific to advertising.
Want to know what this architecture could do for one of your workflows?
