AI Projects — RAG, Agentic AI & Machine Learning

AI projects

Production AI, documented end to end

Every project lists the business problem, the AI approach, the architecture, the evaluation method and the engineering decisions behind it.

RAG

Enterprise RAG Knowledge Assistant

Problem: Internal knowledge was scattered across policy PDFs, wiki pages and ticket history. Teams spent significant time searching, and generic LLM answers could not be trusted because they were ungrounded.
Solution: A retrieval pipeline that ingests and normalises documents, embeds semantically coherent chunks into a vector store, retrieves and reranks candidates, then synthesises an answer constrained to retrieved context with inline citations.
AI approach: Hybrid retrieval (dense embeddings + keyword) with a cross-encoder reranker, structure-aware chunking, and a strict grounding prompt that must abstain when retrieved context is insufficient.

Architecture

Documents → Chunking → Embeddings → Vector Database → Retrieval → Reranking → LLM → Response

Key capabilities

Citation-backed answers
Abstains when context is insufficient
Metadata filtering by document type and access scope

PythonLangChainFAISSFastAPIRedisDockerAWS

Impact · Turned unstructured internal documentation into a trustworthy, auditable answering surface where every response can be traced back to its source passage.

Read the full case study

Agentic AI

Agentic Workflow Orchestrator

Problem: Operational requests required chaining several systems together manually — lookups, validations, record updates and notifications — which was slow and inconsistent.
Solution: A planner-executor agent with a typed tool registry, deterministic validators between steps, retry and fallback policies, and a full execution trace for every run.
AI approach: Task decomposition into a plan, tool selection with schema-validated arguments, execution with bounded retries, and a validator gate before any write operation.

Architecture

User → Agent → Planner → Tools → APIs → Knowledge → LLM → Validator → Action

Key capabilities

Task planning and decomposition
Schema-validated tool calling
Deterministic validation gates

PythonLangChainFastAPIREST APIsRedisDocker

Impact · Replaced brittle manual chains with an auditable, policy-bounded automation layer that fails safely instead of guessing.

Read the full case study

Machine Learning

Customer Churn Prediction Service

Problem: Retention teams reacted to churn after it happened, with no ranked view of which customers were at risk or why.
Solution: A reproducible feature pipeline feeding a gradient-boosted classifier, packaged behind a low-latency scoring API with drift monitoring and scheduled retraining.
AI approach: Behavioural and transactional feature engineering, class-imbalance handling, probability calibration and feature-attribution explanations returned alongside each score.

Architecture

Data → Feature Engineering → Model Training → Evaluation → Model Registry → API → Monitoring

Key capabilities

Calibrated churn probability
Per-prediction feature attribution
Versioned features and models

Pythonscikit-learnXGBoostSQLAirflowFastAPIDockerAzure ML

Impact · Gave retention teams a ranked, explainable risk list they can act on before a customer leaves, rather than a post-mortem report.

Read the full case study

NLP

Document Intelligence & Entity Extraction

Problem: High volumes of unstructured documents were manually triaged and keyed into core systems, creating backlog and transcription errors.
Solution: A pipeline combining layout-aware text extraction, a classifier for document routing, and LLM-based structured extraction validated against a strict schema.
AI approach: Text and layout extraction, document-type classification, schema-constrained extraction with per-field confidence, and human review routing below threshold.

Architecture

Ingest → Text & Layout Extraction → Classification → Structured Extraction → Validation → Downstream System

Key capabilities

Document type routing
Schema-validated field extraction
Per-field confidence scoring

PythonPyTorchFastAPIKafkaSQLAWS

Impact · Shifted document handling from full manual keying to exception-only review, with traceability from every extracted field back to the source document.

Read the full case study

Recommendation Systems

Real-Time Personalisation Engine

Problem: Content and offers were delivered from static segments that ignored what the customer was doing in the current session.
Solution: A hybrid retrieval-and-ranking recommender: candidate generation from collaborative and content signals, ranked by a learned model with session context and business rules applied last.
AI approach: Two-stage recommendation with embedding-based candidate retrieval, a ranking model over session and profile features, and a deterministic business-rule layer.

Architecture

Events → Feature Store → Candidate Retrieval → Ranking Model → Business Rules → Delivery API

Key capabilities

In-session context awareness
Cold-start handling via content signals
Business rule overlays

PythonKafkaRedisSparkFastAPIAWS

Impact · Replaced static segmentation with live, explainable ranking decisions that respond to what the customer is doing right now.

Read the full case study

Predictive Analytics

Transaction Anomaly Detection

Problem: Rule-only monitoring caught known patterns and missed novel ones, while generating high alert noise for analysts.
Solution: An unsupervised anomaly model over engineered behavioural features, combined with existing rules and a scoring layer that ranks alerts by severity.
AI approach: Isolation-forest style unsupervised scoring plus per-entity baselines, blended with deterministic rules and a feedback loop from analyst dispositions.

Architecture

Event Stream → Feature Engineering → Anomaly Scoring → Rule Blend → Alert Ranking → Analyst Feedback

Key capabilities

Near real-time scoring
Per-entity behavioural baselines
Severity-ranked alert queue

Pythonscikit-learnKafkaSparkSQLAirflow

Impact · Extended monitoring beyond known rules while keeping the alert queue at a volume the team can genuinely review.

Read the full case study

Enterprise AI

Enterprise AI Integration Layer

Problem: AI capabilities existed in isolated prototypes with no safe, consistent way for enterprise platforms to consume them.
Solution: A FastAPI gateway in front of model providers and internal AI services, with policy middleware, prompt/response logging, PII redaction and per-consumer quotas.
AI approach: Contract-first API design, provider abstraction, centralised guardrails, and observability on tokens, latency and failure modes.

Architecture

Enterprise App → AI Gateway → Policy & Guardrails → AI Services → Model Providers → Observability

Key capabilities

Single governed AI entry point
PII redaction and content guardrails
Per-consumer quotas and cost attribution

PythonFastAPIREST APIsMicroservicesDockerAzureSalesforce

Impact · Made AI a governed, reusable enterprise capability instead of a set of ungoverned point solutions.

Read the full case study

Generative AI

Streaming Support Copilot

Problem: Agents were switching between knowledge bases and case history mid-conversation, slowing responses and producing inconsistent wording.
Solution: A streaming assistant that combines live conversation context with retrieval over knowledge and case history, producing suggested replies the agent can accept or edit.
AI approach: Rolling conversation summarisation, retrieval over knowledge plus prior resolutions, and token streaming so drafts appear as the conversation develops.

Architecture

Conversation → Context Builder → Retrieval → LLM Stream → Agent UI → Feedback

Key capabilities

Token-streamed suggestions
Grounded in knowledge and prior resolutions
Agent accept / edit / reject feedback

PythonFastAPILangChainFAISSRedisAWS

Impact · Kept agents in one workflow with consistent, source-grounded wording, while every suggestion stays under human control.

Read the full case study

Real-Time AI

Real-Time Event Scoring Pipeline

Problem: Model scores were produced in nightly batches, so decisions were made against stale customer state.
Solution: A Kafka-based streaming pipeline with an online feature store, model scoring workers and a decision topic consumed by downstream platforms.
AI approach: Stream processing with online feature lookup, warm model workers, backpressure handling and replay capability.

Architecture

Event Stream → Online Features → Scoring Workers → Decision Topic → Consumers → Monitoring

Key capabilities

Event-time scoring
Online feature enrichment
Replay from offset for backfill

PythonKafkaRedisSparkDockerAWS

Impact · Moved decisioning from overnight batch to event time, so downstream systems act on current customer state.

Read the full case study