fix(pdf): bug-report sweep — content-TOC + multi-col + bullets + file-check + root-cause cleanup

.~lock.Vasyl-Personal-Projects.pdf#

@@ -0,0 +1 @@
+,blissful-funny-gauss,claude,14.05.2026 16:12,file:///sessions/blissful-funny-gauss/.config/libreoffice/4;

.~lock.Vasyl-Resume.pdf#

@@ -0,0 +1 @@
+,blissful-funny-gauss,claude,14.05.2026 16:12,file:///sessions/blissful-funny-gauss/.config/libreoffice/4;

.~lock.textstack-gemma4-submission-package.pdf#

@@ -0,0 +1 @@
+,blissful-funny-gauss,claude,08.05.2026 15:15,file:///sessions/blissful-funny-gauss/.config/libreoffice/4;

AI-ENGINEER-ROADMAP.md

@@ -0,0 +1,388 @@
+# AI Engineer — Career Roadmap
+
+**Owner:** Vasyl Vdovychenko
+**Created:** 2026-05-13
+**Goal:** Position as AI Engineer and start interviewing in ~6 weeks
+
+---
+
+## Part 1 — Current Profile Audit
+
+### Claims I removed from LinkedIn About (and why)
+
+| Removed | Reason |
+|---------|--------|
+| **"Hybrid search"** (RAG bullet) | Hybrid = vector + BM25 / keyword. My `rag-chatbot-dotnet` uses vector retrieval only. |
+| **"Multi-agent workflows"** (Agent bullet) | No 2+ agents collaborating in my repos. Have single agents with tools, not multi-agent. |
+| **"Response distillation"** (Production AI infra) | Distillation = training a smaller model on outputs of a larger one. I do model **selection** (E4B → E2B), which is different. |
+| **"Kubernetes"** (Full-stack) | Use Docker Compose + Cloudflare Tunnel, not K8s. No K8s manifests in any repo. |
+| **"Anthropic Claude"** (LLM integration) | `.env.example` has `CLAUDE_API_KEY` placeholder, but I haven't verified actual Claude SDK calls in code. **Don't claim until verified.** |
+
+### Final defensible LinkedIn About (every word backed by code)
+
+```
+AI Engineer with 10+ years in software engineering. Building production AI systems —
+RAG pipelines, agent architectures, LLM orchestration, and observability for ML workloads.
+
+Currently a Senior Software Engineer at Pinnacle focused on SDK and testing tooling;
+pursuing AI engineering through personal open-source work. Looking to make this my
+full-time focus.
+
+What I build:
+
+• RAG pipelines — Pinecone vector retrieval, OpenAI embeddings, document chunking,
+  context-grounded generation via Semantic Kernel
+
+• Agent architectures — ReAct loops, tool orchestration, MCP server implementation,
+  function calling, single-agent RAG workflows
+
+• LLM integration — OpenAI, Ollama, Semantic Kernel; model selection,
+  quantization tradeoffs, prompt design
+
+• Production AI infrastructure — cost optimization ($0.002/load test cycle),
+  multi-tier caching, fall-back routing, model selection
+
+• Observability for ML systems — distributed tracing (OpenTelemetry + Aspire),
+  load testing at burst (63K req, p95 20.5ms validated)
+
+• Full-stack delivery — ASP.NET Core, React, React Native (Expo), Postgres,
+  Docker Compose, CI/CD
+
+Open to: AI Engineer · ML Engineer · LLM Engineer · AI Infrastructure Engineer ·
+AI Platform Engineer · Applied AI Engineer.
+
+Background: 10+ years building scalable cloud systems on .NET / Azure stack.
+```
+
+### Headline (final)
+
+```
+AI Engineer | RAG · Agents · LLM Infrastructure | 10+ years in software engineering
+```
+
+---
+
+## Part 2 — AI Engineer Skill Matrix
+
+What an "ideal AI Engineer" portfolio typically has, and where I stand.
+
+Legend: ✅ have it · ⚠️ partial · ❌ missing
+
+### LLM Integration
+
+| Skill | Status | Evidence |
+|-------|--------|----------|
+| OpenAI API | ✅ | rag-chatbot, textstack, AiAgents |
+| Anthropic Claude API | ⚠️ | .env exists, real SDK call to verify |
+| Google Gemini | ⚠️ | .env exists, real SDK call to verify |
+| Ollama (open-source models) | ✅ | textstack (Gemma 4) |
+| Multi-provider abstraction | ✅ | AI_PROVIDER switch in AiAgents |
+
+### RAG (Retrieval-Augmented Generation)
+
+| Skill | Status | Evidence |
+|-------|--------|----------|
+| Vector databases (Pinecone/Qdrant/Chroma) | ✅ | Pinecone in rag-chatbot |
+| OpenAI embeddings | ✅ | rag-chatbot |
+| Document chunking | ✅ | rag-chatbot (Wikipedia indexer) |
+| Hybrid search (vector + BM25) | ❌ | — |
+| Re-ranking (Cohere, etc.) | ❌ | — |
+| Long-context strategies | ❌ | — |
+| RAG evaluation (RAGAS) | ❌ | — |
+
+### Agents
+
+| Skill | Status | Evidence |
+|-------|--------|----------|
+| Function calling | ✅ | FunctionRegistry.cs in AiAgents |
+| Tool orchestration | ✅ | ConsoleAgent (weather tool) |
+| ReAct loop | ✅ | AiAgents docs + code |
+| Agent memory | ✅ | Memory systems in AiAgents docs |
+| **MCP server** | ✅ | **Real McpServer (tools + prompts + resources)** |
+| MCP client | ❌ | Built server only, not client |
+| Multi-agent collaboration | ❌ | — |
+| Agent evaluation | ❌ | — |
+
+### Local LLM / On-device
+
+| Skill | Status | Evidence |
+|-------|--------|----------|
+| Ollama | ✅ | textstack |
+| WebLLM (browser inference) | ✅ | ReplyMate |
+| Quantization knowledge (Q4, GGUF) | ✅ | VRAM math post |
+| GPU offload tuning | ✅ | VRAM math post (2.5× measured) |
+| llama.cpp directly | ❌ | — |
+| vLLM / TGI / Triton | ❌ | — |
+
+### Production AI Infrastructure
+
+| Skill | Status | Evidence |
+|-------|--------|----------|
+| Caching strategies | ✅ | textstack disk cache + IndexedDB |
+| Cost optimization | ✅ | $0.002/load test (measured) |
+| Fall-back routing | ✅ | MC fallback cascade |
+| Rate limiting | ✅ | textstack rate-limit middleware |
+| Async inference / queues | ✅ | textstack worker |
+| A/B testing models | ❌ | — |
+| Shadow deployment | ❌ | — |
+
+### Observability / MLOps
+
+| Skill | Status | Evidence |
+|-------|--------|----------|
+| Distributed tracing OTel | ✅ | xUnitOTel + textstack |
+| Load testing | ✅ | LoadSurge (63K req validated) |
+| Latency budgets | ✅ | Load test report |
+| Cost / token tracking | ⚠️ | Partial, not explicit dashboard |
+| LLM eval suites | ❌ | — |
+| LangSmith / Helicone | ❌ | — |
+| Prompt versioning | ❌ | — |
+
+### ML Fundamentals
+
+| Skill | Status | Evidence |
+|-------|--------|----------|
+| Embeddings | ✅ | rag-chatbot |
+| Transformer arch (engineer-level) | ⚠️ | NVIDIA course in progress |
+| Quantization formats (GGUF/AWQ) | ⚠️ | Knows via Ollama, not deep |
+| Tokenizers (BPE) | ⚠️ | Concept, not hands-on |
+| Fine-tuning / LoRA / PEFT | ❌ | — |
+| RLHF / DPO | ❌ | — |
+
+### Frameworks
+
+| Skill | Status | Evidence |
+|-------|--------|----------|
+| Semantic Kernel | ✅ | rag-chatbot |
+| Microsoft.Extensions.AI | ✅ | rag-chatbot |
+| LangChain | ❌ | — |
+| LlamaIndex | ❌ | — |
+| DSPy | ❌ | — |
+| AutoGen / CrewAI | ❌ | — |
+| Haystack | ❌ | — |
+
+### Deployment
+
+| Skill | Status | Evidence |
+|-------|--------|----------|
+| Docker / Docker Compose | ✅ | every repo |
+| Azure (cloud) | ✅ | Pinnacle background |
+| AWS / GCP | ⚠️ | Familiar, not visible in repos |
+| Kubernetes | ❌ | use compose |
+| GPU cluster mgmt | ❌ | — |
+| Model serving (vLLM, Triton) | ❌ | — |
+
+### Safety / Eval
+
+| Skill | Status | Evidence |
+|-------|--------|----------|
+| Prompt injection defense | ⚠️ | textstack SeoPromptSanitizer (basic) |
+| Guardrails | ❌ | — |
+| Red-teaming | ❌ | — |
+| RAGAS / eval frameworks | ❌ | — |
+
+### Languages
+
+| Lang | Status | Notes |
+|------|--------|-------|
+| C# / .NET | ✅✅✅ | 12 years |
+| TypeScript | ✅ | textstack frontend, ReplyMate |
+| **Python** | ❌ | **Biggest gap for AI roles** |
+| SQL | ✅ | Postgres |
+
+---
+
+## Part 3 — 6-Week Learning Plan
+
+**Goal:** Close the top 3 most-demanded gaps — Python, LangChain/LlamaIndex, RAG evaluation — before flipping LinkedIn Open-to-Work toggle.
+
+### Cadence assumption
+
+Day job: ~40h/week. Available learning time: ~8-10h/week (evenings + weekends).
+
+Each week has one **shippable artifact** (a repo, doc, or LinkedIn-ready milestone).
+
+---
+
+### Week 1 (May 13 – May 19) — Python AI Foundations
+
+**Goal:** Stop being a "C# guy who knows AI" — establish Python AI workflow.
+
+**Tasks:**
+- [ ] Set up Python 3.12 + uv + ruff + ipython on dev machine
+- [ ] Install OpenAI Python SDK, run hello-world chat completion
+- [ ] Install Anthropic Python SDK, run hello-world (covers the Claude gap honestly)
+- [ ] Re-implement `rca` (your LLM-powered test-log analyzer) in Python — single file, ~200 LoC
+- [ ] Add it to GitHub as `rca-py` with a clear README
+
+**Deliverable:** `github.com/mrviduus/rca-py` — Python port with both OpenAI and Claude support.
+
+**Why:** This single repo lets me legitimately add "Python · Anthropic Claude" to my LinkedIn skills.
+
+**Resources:**
+- Anthropic SDK quickstart: https://docs.anthropic.com/en/api/getting-started
+- OpenAI Python SDK: https://github.com/openai/openai-python
+
+---
+
+### Week 2 (May 20 – May 26) — LangChain RAG
+
+**Goal:** Build the same thing my `rag-chatbot-dotnet` does, but in LangChain. This is the most-searched framework on LinkedIn.
+
+**Tasks:**
+- [ ] Read LangChain "Get started" + "RAG tutorial"
+- [ ] Build a RAG chatbot in Python + LangChain that mirrors my rag-chatbot-dotnet (Wikipedia → Pinecone → OpenAI)
+- [ ] Add semantic chunking (`langchain.text_splitter.RecursiveCharacterTextSplitter` with overlap)
+- [ ] Add streaming responses
+- [ ] Document the architecture in README with a diagram
+
+**Deliverable:** `github.com/mrviduus/rag-chatbot-langchain` — Python equivalent with proper chunking strategy documented.
+
+**Why:** Recruiter searches for "LangChain" return 10× more results than "Semantic Kernel". This single repo makes me searchable.
+
+**Resources:**
+- LangChain RAG tutorial: https://python.langchain.com/docs/tutorials/rag/
+- Pinecone + LangChain: https://docs.pinecone.io/integrations/langchain
+
+---
+
+### Week 3 (May 27 – Jun 2) — Hybrid Search + Re-ranking
+
+**Goal:** Close the "hybrid search" gap honestly — actually build it.
+
+**Tasks:**
+- [ ] Add BM25 retriever to the LangChain rag-chatbot (from `langchain_community.retrievers`)
+- [ ] Combine BM25 + vector via `EnsembleRetriever` (50/50 weight)
+- [ ] Add Cohere or local re-ranker on top-N results
+- [ ] A/B test: pure vector vs hybrid+rerank on 20 hand-written queries — log win rate
+- [ ] Write up findings as a dev.to post: "Hybrid search measured: I tested vector vs BM25 vs ensemble"
+
+**Deliverable:**
+- Updated rag-chatbot-langchain with hybrid + rerank
+- A dev.to post with measurement table
+
+**Why:** Now "hybrid search" is REAL on my profile. Plus another technical post for portfolio.
+
+**Resources:**
+- LangChain EnsembleRetriever: https://python.langchain.com/docs/how_to/ensemble_retriever/
+- Cohere rerank: https://docs.cohere.com/docs/reranking
+
+---
+
+### Week 4 (Jun 3 – Jun 9) — RAG Evaluation with RAGAS
+
+**Goal:** Add LLM-eval skill — a senior-marker most engineers don't have.
+
+**Tasks:**
+- [ ] Install RAGAS (`pip install ragas`)
+- [ ] Build a 30-question eval dataset for the rag-chatbot (manual, ~2 hours)
+- [ ] Run RAGAS metrics: faithfulness, answer relevancy, context precision, context recall
+- [ ] Generate a metrics report (markdown table)
+- [ ] Optionally: add a CI workflow (`.github/workflows/ragas-eval.yml`) that runs on every PR
+
+**Deliverable:**
+- RAGAS eval suite in rag-chatbot-langchain repo
+- README section showing the metrics
+- (stretch) GitHub Actions running RAGAS on every commit
+
+**Why:** "LLM evaluation" with concrete metrics is what separates senior from mid AI engineer candidates.
+
+**Resources:**
+- RAGAS: https://docs.ragas.io/en/stable/
+- LangSmith for tracing (optional): https://docs.smith.langchain.com/
+
+---
+
+### Week 5 (Jun 10 – Jun 16) — Multi-Agent Collaboration
+
+**Goal:** Close the "multi-agent" gap honestly — build it.
+
+**Tasks:**
+- [ ] Pick CrewAI or AutoGen (CrewAI is friendlier; AutoGen is Microsoft so synergies with my .NET background)
+- [ ] Build a 3-agent system: **Planner → Researcher → Critic**
+  - Planner breaks down a question into sub-tasks
+  - Researcher uses RAG (your existing LangChain pipeline) to gather facts
+  - Critic reviews the synthesis and asks for clarifications
+- [ ] Pick a non-trivial task: "Plan a 3-day technical conference agenda from a PDF of past schedules"
+- [ ] Document the agent topology with a diagram
+
+**Deliverable:** `github.com/mrviduus/multi-agent-research` — working 3-agent pipeline
+
+**Why:** Now "multi-agent workflows" is REAL. Also opens conversation in interviews about agent design tradeoffs.
+
+**Resources:**
+- CrewAI: https://docs.crewai.com/
+- AutoGen: https://microsoft.github.io/autogen/
+
+---
+
+### Week 6 (Jun 17 – Jun 23) — Polish + LinkedIn Flip
+
+**Goal:** Update profile to reflect new skills and start interviewing.
+
+**Tasks:**
+- [ ] Update LinkedIn About to add new bullets:
+  - Python in LLM Integration
+  - Hybrid search + RAGAS in RAG Pipelines
+  - Multi-agent in Agent Architectures
+- [ ] Add Featured items: rca-py, rag-chatbot-langchain, multi-agent-research
+- [ ] Add Skills: Python, LangChain, LlamaIndex (if used), RAGAS, CrewAI / AutoGen
+- [ ] Reorder Top Skills: LLM, RAG, AI Engineering, **Python**, Agents
+- [ ] **Flip "Open to Work" toggle ON** (Recruiters only visibility)
+- [ ] Add 5 AI Engineer job titles in Open To Work
+- [ ] Write a LinkedIn post: "What I built in 6 weeks pivoting to AI Engineering" — links to all 4 new repos
+- [ ] Connect with 10-15 AI engineers in Toronto / Waterloo / remote-Canada
+
+**Deliverable:** Profile that reads as senior AI Engineer with proof, recruiter messages start arriving within 1-2 weeks.
+
+---
+
+## Part 4 — Top 3 Gaps to Close (TL;DR)
+
+In priority order:
+
+1. **Python** — Week 1. Without this, "AI Engineer" doesn't read as believable for most roles.
+2. **LangChain or LlamaIndex** — Week 2. Recruiter searches for these in 10× more volume than Semantic Kernel.
+3. **LLM evaluation (RAGAS)** — Week 4. Senior-marker that most candidates miss.
+
+Everything else (multi-agent, hybrid search, rerank) is layered on top across weeks 3 and 5.
+
+---
+
+## Part 5 — Nice-to-Have (After Week 6)
+
+If interview pipeline is slow or I want extra ammo:
+
+| Skill | Why | Time |
+|-------|-----|------|
+| Real Anthropic Claude integration code | Closes the .env-only gap | 1 evening |
+| Vector DB diversity (add Qdrant or Chroma) | Shows you can pick, not just use one | 1 weekend |
+| Fine-tuning experience (LoRA) | One HuggingFace training run = resume signal | 1 weekend |
+| Kubernetes basics (deploy textstack to k3s) | Required for some enterprise roles | 1 week |
+| LangSmith / Helicone observability | LLM-specific tracing tools | 1 evening |
+
+---
+
+## Part 6 — Success Metrics
+
+How I know this worked:
+
+- [ ] All 4 new repos live on GitHub with README + topics
+- [ ] LinkedIn About updated with backed claims (every word defensible)
+- [ ] Open To Work flipped on (recruiters only)
+- [ ] LinkedIn search for "AI Engineer Canada" surfaces my profile in top 50
+- [ ] First recruiter inreach within 2 weeks of Open To Work flip
+- [ ] First technical screen by week 8 (2 weeks after flip)
+
+---
+
+## Decision Log
+
+| Date | Decision | Reason |
+|------|----------|--------|
+| 2026-05-13 | Skip $125 NVIDIA cert payment | Senior engineers don't need "Fundamentals" cert; portfolio speaks louder |
+| 2026-05-13 | Headline: "AI Engineer \| RAG · Agents · LLM Infrastructure \| 10+ years in software engineering" | Capability-focused, no project URL, no "Local LLM" niche |
+| 2026-05-13 | About: capability-led, not project-led | Senior engineer About reads as skills, not portfolio site |
+| 2026-05-13 | Don't flip Open to Work yet | 1 month buffer to add Python + LangChain + RAGAS + multi-agent before recruiter spam |
+| 2026-05-13 | Add textstack as separate Experience entry (Self-employed, AI Engineer) | Recruiter search filters by Experience title — needs that entry to be findable |
+| 2026-05-13 | "Share profile updates" already OFF | Profile changes don't broadcast to colleagues |