1. Capability Overview Rate yourself first (pick a level in the checklist below); the radar and bars update automatically. Attainment = current level ÷ target level.
53Total skills
0At or above target
0%Overall attainment
BusinessWeakest domain
Attainment by Domain
Outer ring = target (100%)
2. Skill Mind Map Click a domain to expand its skills; click any node to see its requirements, self-check, and set your current level on the right. Node fill tracks your level.
L0 No ExposureL1 Aware of the ConceptL2 Can Do with GuidanceL3 Can Own It and Explain ItL4 Can Set the Standard / Teach ItBorder = parent domain
3. The 53-Skill Checklist The fastest place to rate: choose a level in the dropdown on each row. Green ✓ = at or above target. Ratings are stored in this browser and persist.
Changes apply instantly and are stored locally. Clearing browser data loses them — export a JSON backup.
The table is wide — swipe sideways to see every column →
| Skill | Target | Requirements | Self-check (answerable = at target) | Current |
|---|---|---|---|---|
| Business Discovery — Turning a client's "we want to do AI" into a concrete, measurable use case. | ||||
| Discovery Interviews | L3 | Using structured questions to surface the business process, data volume, current systems, decision chain and success criteria within 30–60 minutes | Given 30 minutes with a business owner, can you come away with a concrete number for what counts as success? | |
| Business Process Mapping | L3 | Draw the end-to-end current-state process, annotating time spent, owner and error rate at each step | Can you map a client process end to end and name the step that costs the most minutes? | |
| AI Opportunity Identification | L3 | Decide which steps are a fit for AI and which are not — and articulate why not | Can you name three scenarios that look AI-ready but that you would advise against — and why? | |
| Success Metrics Definition | L3 | Translate vague expectations into measurable metrics (accuracy, turnaround time, cost, coverage) | When a client says "improve efficiency", can you turn it into a measurable number on the spot? | |
| ROI & Business Case | L3 | Work out the cost-benefit and make the case in three sentences | Given a set of business figures, can you compute the payback period and show a checkable working? | |
| Pricing & Commercial Models | L2 | Know the pricing models for AI projects (fixed-price / time-and-materials / subscription) and when each applies | Can you explain when an AI module should NOT be priced by person-months? | |
| Domain Knowledge | L3 | Deep working knowledge of at least one vertical — its processes and its regulatory constraints | Can you name three red lines that AI must not cross in your target industry? | |
| Stakeholders & Decision Chain | L2 | Identify who uses it, who decides and who can veto — and prepare material for each | Can you list every key role on a project and what each one actually cares about? | |
| AI Engineering — From model calls to inference in production: designing and building the AI application itself — the technical core of an AI-FDE. | ||||
| LLM Capabilities & Limitations | L3 | Understand tokens, context windows and probabilistic generation; know where hallucinations come from and the four scenarios where the model will fail | Can you list three scenarios where the model is guaranteed to fail, and explain why? | |
| Prompt Engineering | L3 | The four elements of a structured prompt; prompt versioning and regression testing | After changing a prompt, can you say exactly what improved and by how much? | |
| Context Engineering | L3 | Context budgeting, compression and summarization, retrieval injection, tool-result trimming, long-session strategy | Can you write a token budget table and say which part gets cut first when you go over? | |
| End-to-End RAG | L3 | Parsing → chunking → embedding → retrieval → rerank → generation → citation tracing | When the knowledge base answers wrong, can you isolate which stage of the pipeline is at fault? | |
| Document Parsing | L3 | OCR on scanned pages, table reconstruction, multi-column layouts, heading hierarchy recovery | Given a PDF with scanned pages and merged cells, can you parse it into something usable? | |
| Tool Calling | L3 | Tool design principles, parameter and return schemas, and descriptions written for the model to read | Would another person reading your tool description know when to call it? | |
| MCP Development | L3 | Build an MCP server independently (tools, resources, transport, auth) | Can you wrap an internal query endpoint as an MCP server and get a client to call it? | |
| AI Engineering in Python | L3 | LLM SDK calls, streaming, async concurrency, timeouts and retries, error handling; can turn AI logic into maintainable code | Given a requirement, can you write the whole path in Python — call → parse → validate → retry → persist? | |
| Agent Design & Implementation | L3 | Agent loop, state, memory, planning, termination conditions, error recovery | Can you draw the agent loop and mark where each step can fail? | |
| Multi-Agent: When & When Not | L2 | Know when NOT to go multi-agent, and why lead-worker handoff beats free-form group chat | Can you give three reasons not to go multi-agent? | |
| AI Coding Agents | L2 | Understand the coding agent core loop; can run it and use it on real projects | Can you get a coding agent to complete a small real requirement end to end? | |
| Evaluation (Evals) | L3 | Building eval sets (five question types), the four RAG metrics, agent metrics, regression testing | After a change, can you report which number moved from what to what? | |
| Four Layers of Guardrails | L3 | Defense in depth across the input, model, tool and output layers | Can you name every check a user's sentence passes through before it reaches the model? | |
| AI Gateway | L3 | A single entry point: multi-model routing, auth, rate limiting, quotas, audit logs, cost attribution | Can you build one entry point that makes every AI call controllable, traceable and billable? | |
| On-Premise Inference | L2 | vLLM / Ollama-class stacks; can stand up an open-weight model inside the client's network | If the client requires a fully internal network, can you get a usable model running within a week? | |
| Quantization & Capacity Planning | L2 | Choosing a quantization scheme, the relationship between VRAM and concurrency, capacity planning | When the client asks "how many GPUs do we need", can you answer with a defensible number? | |
| Software Engineering — The engineering foundation that puts AI capabilities into real software. | ||||
| Python | L3 | Language fundamentals, async, data handling, packaging; can read and modify other people's code (AI-side SDK work belongs to the AI Engineering domain) | Given an unfamiliar piece of Python, can you read it and reshape it into what you need? | |
| Java / Enterprise Backend | L2 | Can integrate and modify an existing enterprise backend | Can you add an endpoint to a Java project and get the tests passing? | |
| REST API Design & Integration | L3 | Interface contracts, error handling, idempotency, timeouts and retries | Can you design an endpoint an agent can call without taking the system down? | |
| Databases & SQL | L3 | Schema design, query optimization, access control, injection protection | Can you build a query tool that only returns the caller's own department data? | |
| Structured Output | L3 | Get the model to reliably emit JSON / enums / tables, with validation and retry | When the model returns a malformed response, does your program self-correct or crash? | |
| Git & Version Control | L2 | Branching, diffs, rollback, collaboration conventions | When AI-generated code breaks something, can you roll back cleanly? | |
| Code Comprehension & Refactoring | L3 | Read an unfamiliar codebase, locate the change point, assess blast radius | Given a project you've never seen, can you find what to change within an hour? | |
| Low-Code Platforms (Dify, etc.) | L3 | Build a working application on a low-code platform fast, and manage its configuration like an engineer | Can you build a knowledge base in Dify within half a day — with citations and refusal to answer? | |
| Enterprise Integration — Wiring AI into existing systems, within the enterprise's access control and audit requirements. | ||||
| Authentication & Identity | L3 | SSO, tokens, service identities, and attributing AI-agent actions to a real identity | For an action initiated by AI, can you trace it back to a specific person's identity? | |
| Authorization & RBAC | L3 | Role-based permission models, least privilege, and protection against privilege escalation | Can you prove this agent cannot reach data it should not see? | |
| Audit & Traceability | L3 | Action logs, tamper-evident records, replayability | If the client wants to replay an AI action from three months ago, can you pull it up? | |
| Enterprise System Integration | L3 | How to integrate with ERP / CRM / OA / core business systems — and where it bites | If the client only has legacy systems with no API, can you offer three integration paths? | |
| Rate Limiting & Quotas | L2 | Concurrency control, budget circuit breakers, per-department quotas | If a department blows through its budget, does your system stop automatically? | |
| Secrets & Configuration Management | L2 | Key rotation, environment isolation, and keeping configuration out of the repo | Is there any chance someone could dig your keys out of the repository? | |
| Data Security & PII Redaction | L3 | PII detection and redaction, data classification, data-residency constraints | Can you guarantee client-sensitive data never lands in the model vendor's logs? | |
| Operations Engineering — Keeping AI applications stable, observable and cost-accountable in production. | ||||
| Linux & Service Deployment | L3 | Processes, logs, networking, resources, running it as a service | When a service goes down, can you find the cause within three minutes without reading docs? | |
| Docker / Compose | L3 | Images, orchestration, networking, volumes, health checks, one-command start/stop | On a different machine, can you bring the whole stack up with one command? | |
| Observability | L3 | Six things made visible: traces, tool calls, cost, latency, logs, metrics | When the client asks "what did the AI actually do", can you pull up the full trace? | |
| Cost Engineering | L3 | Token cost models, cache reuse, tiered model routing, unit economics | When the client asks "what does a year cost", can you work it out on the spot? | |
| Incident Recovery | L2 | Degradation paths, retry and circuit breaking, recovery time objectives | When the model service is unavailable, does your app crash or degrade? | |
| Monitoring & Alerting | L2 | Thresholds on key metrics, alert rules, on-call response | If quality quietly degrades, will you know before the client does? | |
| Delivery & Communication — Turning technology into value the client recognises — where an FDE closes the deal. | ||||
| Rapid POC Delivery | L3 | A standard playbook for going from discovery to a demonstrable POC in five days | Given a vague requirement, can you produce something runnable in five days? | |
| Demos & Storytelling | L3 | A three-minute script: pain → real data → before/after → failure boundaries → value → next step | After you present, can the client repeat your value proposition back to you? | |
| Technical Documentation | L3 | The nine artifacts: problem, discovery, solution, architecture, implementation, evaluation, deployment, demo, value | Could someone else reproduce your project from your documentation? | |
| Client Communication & Expectation Management | L3 | Manage expectations, state the boundaries, decline unreasonable requests | When the client asks for something impossible, can you say no without losing the deal? | |
| Compliance & AI Governance | L3 | Generative AI regulations, algorithm filing, data residency, content safety, audit retention | When the client asks about compliance, can you answer with three or more points instead of deferring to legal? | |
| Failure Boundaries & Fallbacks | L3 | Proactively define failure scenarios, human takeover, rollback, idempotency | Does your proposal state clearly what happens when the AI gets it wrong? | |
| Training & Knowledge Transfer | L2 | Transfer capability to the client team instead of keeping yourself on the critical path | Three months after delivery, can the client run it without you? | |
| Delivery Cadence & Project Management | L3 | Scope cutting, milestones, surfacing risks early | When a deadline will slip, can you say so a week ahead and bring options? | |
4. Levels & Thresholds The most common self-assessment mistake is treating "I have used it" as "I can do it." So every level has a testable standard.
| Level | Name | Standard | Typical evidence |
|---|---|---|---|
| L0 | No Exposure | Haven't heard of it, or only seen the term | — |
| L1 | Aware of the Concept | Can say what it is and what problem it solves, but has not done it | Can explain the concept to someone else |
| L2 | Can Do with Guidance | Can follow docs or an example to get it running; needs help when something breaks | Has a working demo |
| L3 | Can Own It and Explain It | Can do it from scratch unaided, handle common failures, and explain it to a client in two minutes | Real project output with quantified results |
| L4 | Can Set the Standard / Teach It | Can define team conventions and selection criteria, teach others, and improve other people’s designs | Has a reused standard document, or has grown people |
How the bar is set: skills marked L3 are core (without them you are not an AI-FDE); L2 skills are supporting (you can be shallow, but not blank). There are no L1 items — anything worth only "being aware of" did not make the list.
A healthy distribution: 60% or more at L3 and above, and no domain below 50%. If one domain is clearly low, that is usually your bottleneck.
A healthy distribution: 60% or more at L3 and above, and no domain below 50%. If one domain is clearly low, that is usually your bottleneck.
5. Gaps by Background AI-FDE is a hybrid role — nobody arrives with all of it. Four common backgrounds, four different gaps.
Implementation / Delivery
Starting point:Strong on business context and client communication — communication skills usually clear the bar immediately.
Most often missing: AI Engineering (Evaluation, Context Engineering, on-premise inference) and Operations Engineering (observability, cost). Used to delivering against a given design, so you hesitate when the technical approach is yours to define.
Fix: start with Evaluation — it is the only way to turn "feels good" into "the data says so", and it is the strongest evidence you can show a client.
Most often missing: AI Engineering (Evaluation, Context Engineering, on-premise inference) and Operations Engineering (observability, cost). Used to delivering against a given design, so you hesitate when the technical approach is yours to define.
Fix: start with Evaluation — it is the only way to turn "feels good" into "the data says so", and it is the strongest evidence you can show a client.
Pre-Sales / Solution
Starting point:Storytelling and business cases come naturally; discovery and demos are the fastest to pick up.
Most often missing: hands-on depth. You can present a design but cannot build it — which leaves you exposed when engineers push back or clients drill into details.
Fix: get Python, AI engineering in Python, and a low-code platform to at least L2. You do not need to write beautiful code, but you must be able to build it and make it run.
Most often missing: hands-on depth. You can present a design but cannot build it — which leaves you exposed when engineers push back or clients drill into details.
Fix: get Python, AI engineering in Python, and a low-code platform to at least L2. You do not need to write beautiful code, but you must be able to build it and make it run.
Operations / SA
Starting point:Operations Engineering and Enterprise Integration are strengths — Linux, Docker, permissions and auditing usually start high.
Most often missing: business discovery and storytelling. You can build it but cannot sell it, and the client never sees the value — the biggest hurdle for an ops background.
Fix: practise discovery and demos deliberately, as if they were technical tasks. Record yourself running a demo and watch it back; once is usually enough to see the problem.
Most often missing: business discovery and storytelling. You can build it but cannot sell it, and the client never sees the value — the biggest hurdle for an ops background.
Fix: practise discovery and demos deliberately, as if they were technical tasks. Record yourself running a demo and watch it back; once is usually enough to see the problem.
Engineering
Starting point:Software and AI Engineering come fastest; code-heavy skills are rarely a barrier.
The trap you fall into: stack sprawl. You enjoy building your own wheels and chasing new frameworks, and lose sight of business value and delivery cadence.
Fix: force yourself to write all nine artifacts for every project — especially the business-value one. If you cannot write it, the technology has no business value.
The trap you fall into: stack sprawl. You enjoy building your own wheels and chasing new frameworks, and lose sight of business value and delivery cadence.
Fix: force yourself to write all nine artifacts for every project — especially the business-value one. If you cannot write it, the technology has no business value.
The most overlooked item: "Failure Boundaries & Fallbacks" under Delivery & Communication. Most people spend their effort on making the AI get it right, but what the client actually cares about is what happens when it gets it wrong. Whether your proposal spells out the fallback and the human handoff is what separates a demo artist from someone who can deliver.