Forward Deployed Engineer

Topics

• How forward deployed engineering differs from software engineering and consulting
• Product engineering compared with customer-specific implementation, including when customisation becomes technical debt
• Customer requirements, enterprise buying cycles and field feedback into the product
• Customer trust and engineering credibility when working directly with enterprise stakeholders
• Stakeholder interviews, jobs-to-be-done analysis and current-state workflow mapping
• Manual handoffs, system constraints, user roles and root-cause analysis
• Value hypotheses, workflow baselines, adoption measures and technical success measures
• Scope boundaries, pilot acceptance criteria, kill criteria and change-request management
• Choosing between software, rules, machine learning, RAG and agent workflows
• Using a deterministic workflow when it can solve the problem more safely than an agent
• Build-versus-buy decisions, hosted and self-hosted models, data residency, latency, cost and architecture decision records

Applied work

• Frame an incomplete request for an AI support agent before proposing a solution.
• Handle a mentor-led customer simulation with incomplete and occasionally contradictory requirements.
• Produce a problem-framing memo, discovery transcript, stakeholder map, workflow map and problem hypothesis.
• Prepare a value hypothesis, baseline measurement, success scorecard, scope document and pilot acceptance criteria.
• Calculate workflow measures such as cost per completed task, handling time, error rate, escalation rate, human-review rate and intended-user adoption.
• Do not begin coding until the discovery evidence, scope and customer decision criteria are clear.
• Create an architecture diagram, decision record, risk register and build-versus-buy recommendation.

Review Gate 1: Discovery and Architecture. The project proceeds only after the customer problem, success criteria and architecture are approved.

Topics

• Data-source inventories, relational and document data, data contracts, batch and event-driven ingestion, validation, lineage and schema change
• Personally identifiable information, CRM, ERP and ticketing integrations, enterprise APIs, document parsing and metadata design
• Structured outputs, tool calling, prompt versioning, context management, deterministic boundaries and model routing
• Retries, fallbacks, prompt caching, response validation, provider portability, latency and cost measurement
• Document ingestion, dense and keyword retrieval, hybrid search, reranking, query transformation, freshness, citations, access-aware retrieval and evaluation
• Agent state, planning loops, approval flows, memory design, MCP fundamentals, multi-agent trade-offs, audit trails and sandboxed execution
• One primary stack used deeply: Python, FastAPI, PostgreSQL, Docker, GitHub Actions, OpenTelemetry, a major cloud platform and model-provider portability testing

Applied work

• Build a monitored ingestion pipeline connected to two different data sources.
• Document the integration map, data contract, data-quality report and failure-recovery plan.
• Create a full-stack AI workflow with validated outputs and fallback behaviour.
• Deliver a RAG application with citations, document-level access controls, changing content and malformed-input handling.
• Prepare a retrieval benchmark, golden evaluation set, failure analysis and cost-per-query report.
• Build an agent that completes an enterprise API action, such as routing a support-ticket resolution or updating a CRM record, and asks for approval before a high-risk step.
• A simple chat-with-a-PDF demonstration does not qualify as the required retrieval application.

Review Gate 2: Working Pilot. The working pilot must solve the approved workflow. A polished interface alone is not enough.

Topics

• OAuth, OpenID Connect, role-based access, tenant separation, service accounts, enterprise single sign-on and permission-aware tool use
• API gateways, webhooks, queues, rate limits, idempotency, session management and audit logging
• Evaluation-driven development, golden datasets, retrieval measures, generation quality, tool-call correctness, task completion, safety tests and launch thresholds
• Human evaluation, LLM-as-judge limitations, judge calibration, regression tests, online and offline evaluation, and evaluation in CI/CD
• Threat modelling, prompt injection, data exfiltration, insecure tool use, secrets and model supply-chain risk
• Encryption, privacy, data retention, human review, OWASP risks for LLM applications, India's data-protection environment, GDPR fundamentals, sector-specific controls and customer security reviews
• Docker, Kubernetes and serverless deployment, infrastructure as code, environment separation, feature flags, secrets, CI/CD, model and prompt observability, rollback, disaster recovery, VPC and on-premise deployment patterns

Applied work

• Deliver a multi-user AI application with role-based access and auditable tool actions.
• Create an automated evaluation pipeline that blocks releases below the approved quality threshold.
• Produce an evaluation plan, test dataset, baseline report, regression report and launch recommendation.
• Run an independent red-team exercise and document the threat model, findings, remediation log, data-flow diagram and security-review response.
• Deploy with an automated release, tracing, alerting and rollback, supported by infrastructure configuration, an observability dashboard and a runbook.

Review Gate 3: Trust. The system must meet evaluation and security thresholds before deployment. A working AI output is not enough without dependable workflow performance.

Topics

• Service indicators, service objectives, availability, load testing, rate-limit handling, latency reduction, capacity planning and incident management
• Cost per successful task, token economics, caching, model routing, batch processing and graceful degradation
• Pilot cohort selection, onboarding, user resistance, adoption measures, overrides, stakeholder communication and the decision to expand or stop a pilot
• Reusable connectors, deployment templates, evaluation libraries, configuration-driven design and product feedback
• Technical debt, feature prioritisation and turning field learning into reusable product improvements

Applied work

• Respond to a simulated outage, model degradation or sudden cost increase.
• Produce an incident timeline, root-cause analysis and revised reliability plan.
• Run a pilot review with stakeholders and prepare a rollout plan, user guide, adoption dashboard, feedback analysis and recommendation.
• Create a reusable accelerator, product feedback memo, generalisation analysis and technical-debt register.
• Judge the pilot by use in the real workflow. A system that users avoid has not succeeded.
• Turn field learning into reusable improvements instead of leaving permanent one-off customer codebases.

Topics

• Operational handover, support boundaries, documentation, knowledge transfer and post-deployment roadmaps
• Executive communication, technical storytelling, architecture trade-offs and demonstrating business value

Final submission

• A working deployment, architecture, evaluation report and security assessment
• A monitoring dashboard, incident postmortem, adoption evidence, ROI analysis and runbook
• A recorded demonstration and a live defence with a senior engineer, customer or business stakeholder, and deployment reviewer

Review Gate 4: Production Readiness. The deployment must meet the approved operational, evaluation and security standard.

Review Gate 5: Customer Handover and Live Defence. The final panel assesses handover readiness, technical judgment and the ability to explain the business value.