Things I’ve designed and built.
Each card has the key design decision, straight from the repo’s README.
LLM auto-router on Azure AI Foundry
Routes each prompt to the cheapest model that can handle it. 98% routing accuracy on 100 labelled prompts vs 62% for keyword rules, and about 27–29% lower cost than sending everything to gpt-5.4 (measured in the repo; 98% is optimistic until a held-out split is added).
Design: An LLM classifier (gpt-5-mini) picks one of four model tiers; its JSON is validated against policy rules, and invalid output falls back to the strongest model. A 100-prompt golden set drives the tuning loop.
Production-grade RAG
A RAG build taken through seven phases, with a walkthrough, architecture diagram, and demo video.
Design: Hybrid BM25 + vector search (alpha 0.5, top 30), a cross-encoder reranker keeps the best 5, superseded docs are filtered out, and a CI eval gate fails any PR that drops retrieval quality.
EMBER — weather-aware fire triage
Bahasa-first AI fire commander for village brigades during Southeast Asian haze fires: ranks satellite detections by spread risk, with a volunteer app and supervisor console. Submitted; awaiting results.
Design: Agents fuse reports from many phones into one shared incident; a human supervisor approves every high-stakes call, and the WITHDRAW-upwind safety rule is deterministic code, not model output.
LakeShift — PL/SQL to Snowflake, proven
Translates legacy Oracle PL/SQL to Snowflake SQL with NVIDIA Nemotron on Nebius Token Factory, then proves every result row by row. On the recorded run: 6/6 objects proven equal, 5,685 rows compared. Nebius x NVIDIA hackathon; submitted.
Design: Planner → parallel Nemotron Nano translators → a verifier that diffs Oracle vs Snowflake row by row; failures go to a docs-only researcher and a fixer that escalates Nano, then Super, then Ultra until the verifier passes.
Agentic Procurement — agents that negotiate and pay
A buyer agent negotiates with seller agents and pays on PayPal: agentic procurement and payables for small businesses. PayPal AI Hackathon; building, submissions close Nov 12.
Design: Agents negotiate, but money moves through PayPal authorize-then-capture: funds are held on agreement and captured only once the deal is confirmed, so an agent never pays without a settled order.
Fraud policy copilot
What you can build with Snowflake CoCo CLI + Cortex: flagged transaction in, policy-cited risk decision, fraud-ring map, case and native alert trail out.
Design: Streamlit in Snowflake: transaction id → customer context → rule matches → risk score with policy citations, then a case and a native Snowflake alert trail.
Foundry eval gate
A CI eval gate that checks agent traces for faithfulness, tool success, and PII leaks before a change ships.
Design: Deterministic, fixture-first checks on JSON traces that exit non-zero, so GitHub Actions or Azure DevOps can block the merge.
ADO pipeline doctor
Turns failed Azure Pipelines and GitHub Actions logs into structured RCA JSON, with human approval required before any fix.
Design: Logs become structured RCA JSON with an explicit human_approval_required flag, so it’s safe to wire into an agent.
MCP-style ops tools
iac-guard-mcp (Terraform/Bicep diff review), azure-ops-mcp (Key Vault compare, ADO diagnose, dry-run + audit), lakehouse-insights-skill, and kc-mcp.
Design: iac-guard uses Checkov when available and falls back to built-in regex policies, returning structured findings JSON that agents can cite.
Infrastructure as code
Hub-and-spoke Terraform modules for Azure networking, Python Key Vault automation, and multi-CI demos across GitHub Actions, Azure DevOps, and GitLab.
Design: Separate hub and spoke Terraform modules plus an example that wires them together: a foundation to extend, not a full landing zone.
More on GitHub.