AI Engineer
Cognizant Technology Solutions · Feb 2024 – Present
- Enterprise Generative AI and Agentic AI delivery for a Tier-1 Saudi bank architecting, evaluating, securing and operating production LLM systems across financial crime, corporate credit, payments operations, customer service and regulatory compliance.
- Architected six specialised agents Supervisor, Transaction Analysis, KYC, Risk, Compliance and Verification in Python and LangGraph, with agent orchestration, persistent state management, conditional routing, retries and human-approval checkpoints across multi-step investigation workflows.
- Integrated agents with authorised customer, KYC, transaction, fraud-risk and case-management APIs through controlled tool calling, and standardised enterprise tool and data access behind an MCP-compatible tool layer where appropriate, decoupling agent logic from individual backend contracts.
- Engineered an enterprise RAG pipeline over AML/fraud policies and investigation procedures document parsing, semantic chunking, metadata enrichment, embedding generation, hybrid vector + BM25 retrieval and re-ranking so case narratives stayed grounded in approved policy with traceable citations.
- Implemented structured prompting and schema-validated JSON outputs for agent-to-agent communication and generated investigation reports, removing free-text drift between agent handoffs.
- Enforced layered guardrails PII protection, prompt-injection detection, tool permissions, RBAC, grounding validation and human-in-the-loop approval so no investigative action executed outside an analyst's authorised scope.
- Built a model evaluation framework measuring faithfulness, hallucination, retrieval quality, tool-call accuracy, agent task completion and safety, and automated Pytest suites over golden datasets covering agent workflows, APIs, RAG retrieval, tool calls and failure scenarios.
- Instrumented agent tracing and production monitoring for latency, token usage, tool execution, failures and end-to-end investigation throughput, using trace and evaluation evidence to drive prompt, retrieval and model decisions.
- Developed the multi-step credit-review agent workflow in Python, LangChain and LangGraph, sequencing document intake, extraction, policy validation, gap detection and credit-memo drafting with recoverable state.
- Built document-intelligence pipelines to extract structured information from financial statements, credit applications and supporting documents, and integrated customer, facility, KYC, financial and credit data through secured enterprise APIs.
- Kept all financial computation deterministic ratios, trends, exposure calculations and validation checks implemented as Python services rather than LLM arithmetic restricting the model to interpretation, summarisation and narrative.
- Implemented RAG over credit policies, lending guidelines, procedures and product documentation, with prompts engineered for financial-document summarisation, risk-factor identification, missing-document detection and draft credit-memo generation.
- Applied structured outputs, validation rules and human-in-the-loop approval so every generated statement traced back to a source document and credit recommendations and exceptions remained under authorised credit-officer control.
- Created evaluation datasets and automated Pytest suites validating extraction accuracy, RAG quality, hallucination, policy grounding and workflow execution, alongside logging, tracing, retries and exception handling for agent and API workflows.
- Developed an agentic payment-exception investigation workflow in Python and LangGraph, with specialised capabilities for transaction diagnostics, payment-policy retrieval, resolution planning, verification and customer notification.
- Integrated live payment and transaction-status APIs as controlled agent tools, applying JSON-schema validation, permission boundaries, timeouts, retries and structured error handling on every tool call.
- Added deterministic workflow controls and idempotency protection around payment-related actions to prevent duplicate or unintended execution, and designed human-approval gates for financial, sensitive and high-risk remediation.
- Built RAG pipelines over payment procedures, exception codes, operational guidelines and troubleshooting knowledge so root-cause explanations and recommendations cited approved operational documentation.
- Implemented guardrails for PII protection, authorisation, prompt injection, unsupported-action refusal and hallucination prevention, with an enforced allowlist constraining the agent to permitted remediation actions.
- Developed Pytest automation covering agent workflows, APIs, tool calls, failure scenarios and exception handling, and implemented tracing and monitoring for agent decisions, tool execution, API latency, failures and resolution outcomes.
- Built an enterprise RAG-based customer-service copilot in Python, FastAPI and LangChain, with document ingestion, cleaning, chunking, metadata enrichment, embedding and indexing pipelines over banking product, policy and process knowledge.
- Implemented hybrid retrieval combining semantic/vector search, BM25 keyword search, metadata filtering and re-ranking to lift answer accuracy across a large, heterogeneous knowledge base.
- Assembled runtime context dynamically from authorised customer and transaction APIs, conversation history and retrieved knowledge within a managed token budget, so responses were personalised without exposing unauthorised data.
- Engineered prompts for Arabic/English query understanding, grounded response generation, summarisation and resolution guidance, with source citations and grounding validation to suppress unsupported answers.
- Added guardrails for PII masking, sensitive banking information, prompt injection, inappropriate requests and access control, applied at both the retrieval and output layers.
- Built a model evaluation framework covering faithfulness, answer relevance, context precision and recall, hallucination, safety and bilingual response quality, with automated RAG and API testing in Pytest over curated golden question-answer datasets.
- Monitored LLM latency, token consumption, retrieval performance, failed requests and response-quality metrics, feeding findings back into prompt, chunking and retrieval tuning.
- Developed the enterprise RAG pipeline over regulatory documents, internal policies, procedures, compliance standards and control documentation, covering parsing, intelligent chunking, metadata tagging, embeddings, vector indexing and document-version handling.
- Built hybrid retrieval and re-ranking to sustain retrieval accuracy across large regulatory and policy repositories, and integrated regulatory knowledge repositories and internal document systems through secured APIs and connectors.
- Enforced RBAC and metadata-level filtering at retrieval time so users surfaced only documents permitted for their role, giving document-level and tenant isolation inside a shared platform.
- Designed prompt templates for regulatory Q&A, policy comparison, summarisation, obligation extraction and compliance-impact analysis, with citation generation and grounding validation tying every response to authoritative sources.
- Developed guardrails covering PII, prompt injection, hallucination, sensitive information and unsupported legal or compliance conclusions, so the copilot informed decisions rather than adjudicating them.
- Delivered a model evaluation framework measuring retrieval precision and recall, faithfulness, citation correctness, answer relevance, hallucination and policy-grounding accuracy, with Pytest regression suites and golden datasets re-run after every model, prompt, embedding or document change.