Agentic AI Real-World Applications
These projects apply the course patterns in increasing order of autonomy.
1. Evidence-backed technical researcher
Outcome: answer a technical question with claim-level citations and visible uncertainty.
- Steps: clarify → plan queries → retrieve → extract evidence → find gaps → synthesize → verify citations.
- Tools: web/search, document reader, source store.
- Evals: source quality, citation alignment, factual support, coverage, latency.
- Guardrails: treat pages as untrusted, prohibit invented sources, cap searches.
- MVP: answer ten known questions from an allowlisted documentation set.
2. Chart and insight assistant
Outcome: transform a dataset and question into a correct, readable chart plus explanation.
- Pattern: generate code → sandbox execution → semantic/visual checks → reflection.
- Evals: correct aggregation, units, labels, missing-data handling, accessibility, insight accuracy.
- Guardrails: isolated runtime, file limits, no arbitrary network, deterministic data checks.
- MVP: CSV input and three chart families.
3. Read-only SQL analyst
Outcome: answer business questions from an approved database.
- Steps: select schema → draft query → static validation → execute → inspect results → explain.
- Evals: query correctness, result faithfulness, row/timeout compliance, sensitive-column access.
- Guardrails: read-only role, allowlisted schemas, query timeout, row limit, audit log.
- MVP: twenty questions against a small analytics database.
4. Email triage and drafting copilot
Outcome: prioritize mail, identify requested actions, and prepare drafts.
- Tools: read thread, contact context, calendar lookup, draft creation.
- Evals: classification, missed obligations, tone, factual grounding, approval compliance.
- Guardrails: never auto-send initially; confirmation for calendar or external disclosure.
- MVP: label and draft only for a sandbox mailbox.
5. Customer-support resolution agent
Outcome: resolve routine cases under policy and escalate exceptions cleanly.
- Steps: identify intent → retrieve account facts → consult policy → propose action → approve/execute → verify.
- Evals: resolution correctness, policy adherence, escalation recall, customer effort, cost.
- Guardrails: policy engine in code, least-privilege tools, refund limits, human approval above thresholds.
- MVP: order-status and FAQ cases, then one reversible account action.
6. Market-research team
Outcome: produce a decision-ready market brief from independent evidence streams.
- Roles: coordinator, customer researcher, competitor researcher, market-signal analyst, skeptic, synthesizer.
- Evals: evidence diversity, freshness, contradiction handling, recommendation traceability.
- Guardrails: bounded parallel work, explicit source dates, no fabricated numbers, one final owner.
- MVP: compare three competitors for one clearly defined buyer segment.
7. Repository maintenance agent
Outcome: deliver a small, proven code change within an explicit permission envelope.
- Steps: read repository guidance → map change → plan → edit one slice → run checks → review diff → summarize evidence.
- Tools: scoped filesystem, tests, linters, browser, version control read operations.
- Evals: task completion, regression rate, boundary violations, review defects, time/cost.
- Guardrails: no destructive commands or deployment without approval; limit files and iterations.
- Workflow: Spec Map Build Prove Learn.
Choosing a first build
Start with the research agent if learning retrieval and evaluation, the chart/SQL assistant if learning external feedback, or the email copilot if learning tool permissions. Build customer support and multi-agent research only after one-agent traces and evals are reliable.