Job description to validated JSON
Turn messy job descriptions into evidence-backed records, with missing-field handling and repeatable tests.
BUILD SOMETHING YOU CAN EXPLAIN
AI Project Library
Go beyond a tutorial copy. Build a useful system, measure where it fails and turn your decisions into portfolio evidence.
Find your next project ↓20 of 20 projects · Estimates are focused build hours, not guaranteed completion times.
Demand forecasting without future leakage
Build a forecast API, compare it with a seasonal baseline and prove that future information never enters training.
Evidence-backed report drafting
Draft short reports from a controlled evidence pack, checking citations and refusing unsupported conclusions.
Experiment readout and decision notebook
Analyse a fictional onboarding experiment with uncertainty, guardrail metrics and an honest decision memo.
LLM structured-output comparison lab
Compare two model configurations on factual extraction, separating accuracy, malformed output, warm-up time and speed.
Permission-aware RAG knowledge assistant
Build grounded search with access checks, citation evaluation and explicit refusal when evidence is missing.
Approval-bound support workflow agent
Build a bounded tool workflow with approval, idempotency and recovery, using a sandbox instead of real customer actions.
Replayable dataset pipeline for AI
Ingest noisy events into a versioned dataset with contracts, deduplication, quarantine and safe reprocessing.
Model release gates and rollback drill
Promote a tiny model only after quality and operational checks pass, then demonstrate rollback when a canary regresses.
Image intake quality gate
Reject unusable product photos, compare rules with an optional small classifier and test lighting-related failures.
AI learning resource recommender
Recommend useful learning resources from skills and goals, with cold-start handling and transparent ranking evaluation.
Read-only text-to-SQL assistant
Answer questions over a fictional database with validated SQL, read-only permissions and explicit ambiguity handling.
Visual product search with text queries
Search a permitted image catalogue using descriptions and compare cross-modal retrieval with a keyword baseline.
Speech transcription evaluation lab
Compare transcripts with human references, separating word errors, critical terms, noise and processing time.
Hand-gesture presentation controller
Control demo slides with explicit hand gestures, cooldowns and a keyboard fallback, while keeping camera frames local.
Sensitive information redaction workbench
Detect and mask selected personal information in fictional text, with span-level tests and a human-review queue.
Service-metric anomaly alert dashboard
Detect unusual synthetic service metrics and evaluate alert precision, missed incidents and detection delay.
Fine-tuned support ticket classifier
Compare a LoRA-adapted Gemma classifier with a prompt-only baseline, using held-out tickets and explicit resource accounting.
Semantic duplicate detection review queue
Find differently worded duplicate records without merging distinct items that happen to share a topic.
Environmental sound classification explorer
Classify permitted everyday sounds and investigate how background noise, chunking and thresholds change results.
Build one project well.
Pick a scope you can complete, create a simple baseline, freeze a test set and write down what fails. A clear README, reproducible evaluation and thoughtful trade-offs are more useful than a long list of unfinished demos.
These are educational blueprints, not downloadable finished applications or hiring guarantees. Google documentation and GitHub references are labeled separately. API usage and optional cloud infrastructure may incur charges.