I build LLM systems and measure them before I trust them.

Three pipelines sit behind one streaming API: hybrid RAG (BM25 plus dense, fused with RRF), GraphRAG and agentic RAG. All three are scored on the same 60 questions, so the choice of default is a measurement and not a preference.1

• Every pull request is evaluated. CI runs lint, 139 tests, Terraform validation and a live evaluation. If answer quality drops below the threshold the merge is blocked. Deploys need approval and roll back automatically.
• Security was tested, not assumed. Token auth, document-level security trimming and PII masking. Injection through retrieved passages succeeded 1.4% of the time on a plain prompt and 0 of 240 times against the agent.
• Costs are known. OpenTelemetry traces every request. The agent costs about $3.54 per 1,000 requests and takes 12.4 s at the median, against about $2 and 3.9 s for the baseline.

Built with Python, FastAPI, Azure AI Search, Azure OpenAI, Container Apps, Terraform, GitHub Actions with OIDC, OpenTelemetry, pytest

Engineers work across datasheets, Excel parameter lists, requirements and CAD, in two languages and several versions. A wrong or outdated value costs time. So every statement in an answer carries its source, and the system says "not found" when the documents don't contain the answer.2

• Exact labels by construction. Three fictional fuel cell systems are defined once in YAML. Generators turn them into documents and record where each value appears, which gives exact gold sources for 70 questions (25 in German) over 196 documents.
• A judge I checked. Faithfulness is scored claim by claim by a model from a different family than the one answering. Against independent labels the judge agreed on 72 of 75 claims and was never more lenient.
• Found the bottleneck. Under 10 users, retrieval p95 went from 18 s to 4.9 s once I measured that reranking took almost all of the time.
• Operations. Langfuse traces, Prometheus metrics, an audit log, a Redis response cache and a CI evaluation gate.

Built with Python, FastAPI, Qdrant, bge-m3 and bge-reranker-v2-m3, Docling, LiteLLM, Langfuse, Postgres, Redis, Docker

A citizen sends a Telegram bot a photo, a voice note in Tunisian Derja, French or English, a text or a GPS location. Agents estimate the water depth, extract who is in danger, retrieve the official flood protocols and grade the incident from P1 to P4. The ticket goes to a dispatcher board and back to the citizen in their own language.3

• The model can't lower a priority. The last step is deterministic code. The safety rules can only raise a grade, and unclear cases go to a human.
• Protocol retrieval: Hit@3 0.95 and MRR 0.935 with structure-aware chunks on Qdrant.
• Vision: a second prompt version cut depth under-estimation from 55% to 29% of photos, and the error on severe floods from 23 cm to 5 cm.
• Distress extraction: 8 of 8 labelled voice and text cases.

Built with n8n, Telegram, Groq (Qwen vision, gpt-oss, Whisper), Gemini, Qdrant, Python

Before a release, someone has to read the test report, the coverage, the linter, the CVE scan and the secrets scan, then decide. My agent collects all of it and returns one verdict with the reasons.4

• Inputs: tests and coverage, lint (ruff), dependency audit, CVEs (Trivy), secrets (Gitleaks) and SonarQube findings.
• Scans decide, the LLM explains. The optional rationale is written after the verdict and can never change it.
• Two ways in. It runs in GitHub Actions, and it is exposed as an MCP server so the platform's other agents can ask for a verdict.

Built with Python, MCP, GitHub Actions, pytest, Trivy, Gitleaks, SonarQube, Docker. The wider platform uses LangGraph, Kafka and gRPC.

The public prototype sits on an open-source fraud-detection project by Mohit Gupta. My part is the gate, its tests, the CI and the Docker hardening.

Next-Concept Modeling writes text coarse to fine instead of one token at a time, and is reported to be up to 31× faster. I rebuilt a toy version from scratch, validated the official checkpoints against the paper, then ran 8 experiments that push it until it breaks. I wrote my predictions down before each run and graded them afterwards.5