Syket Bhattachergee · Chattogram, Bangladesh 🇧🇩
🟢 Open to AI Engineer roles — remote or relocation
I build LLM systems that hold up in production — RAG pipelines with real retrieval strategy, agents as explicit state machines, and MCP servers that expose scoped tools to Claude and other clients.
Software engineer since 2020. AI engineer since 2025. The pivot wasn't a restart. Everything that makes an LLM feature survive contact with real users — state, latency, retries, error handling, knowing what to do when a call fails — is the same engineering I've been doing for five years. The model is just a new dependency, and it's the least reliable one in the stack.
That's why I lead with evals. I'd rather prove a change improved the system than guess.
flowchart LR
Q[User query] --> R{Route}
R -->|needs knowledge| RAG[Hybrid retrieval<br/>BM25 + dense]
R -->|needs action| AG[LangGraph agent]
RAG --> RR[Re-rank] --> CTX[Context assembly<br/>token budget + citations]
AG --> T[Scoped tools / MCP]
T --> HIL{Human-in-the-loop<br/>checkpoint}
HIL --> CTX
CTX --> LLM[LLM call<br/>structured output + validation]
LLM --> V{Output valid?}
V -->|no| LLM
V -->|yes| UI[Streaming UI<br/>token-by-token + sources]
LLM -.trace.-> EV[Eval harness<br/>RAGAS · LangSmith · regression gate]
EV -.gates deploys.-> LLM
style EV stroke:#A78BFA,stroke-width:2px
style HIL stroke:#A78BFA,stroke-width:2px
The two boxes most demos skip are the two I care about most: the eval gate and the human checkpoint.
AI / LLM
Python LangChain LangGraph LangSmith LlamaIndex Pydantic FastAPI MCP RAGAS
OpenAI · Anthropic Claude · Gemini · Hugging Face
Pinecone Qdrant ChromaDB pgvector
Product surface
TypeScript Next.js React React Native Expo Vercel AI SDK Tailwind GraphQL
Infra
Docker PostgreSQL Prisma Redis Vercel GitHub Actions
What each area means in practice →
| Area | In practice |
|---|---|
| 🔍 RAG | Chunking strategy selection, hybrid BM25 + dense search, cross-encoder re-ranking, metadata filters, citation UX |
| 🤖 Agents | LangGraph state machines, tool-use, supervisor handoffs, parallel sub-agents, human-in-the-loop checkpoints |
| 🧩 Context engineering | Token budgeting, memory strategies, context-window management, prompt caching, model routing for cost |
| 📊 Evals | Golden sets, LLM-as-judge, regression gates in CI, tracing production failures back into the dataset |
| 🔌 MCP | Scoped, secure tool servers for Claude and other clients |
| 💬 Streaming UI | Vercel AI SDK, RSC, generative UI from tool calls, optimistic updates, graceful failure states |
CreoWis Technologies — Associate Software Engineer · Jan 2024 – Present
Built Finanion, a personal finance dashboard. Built a centralized auth system with RBAC (Keycloak-style). Built a JSON→PDF report generation pipeline. Shipped next-i18n-test-automation — 30% faster than writing multi-language test coverage from scratch.
Woofmeets — SDE II · Jan 2022 – Dec 2023 Led frontend from zero on a pet services platform. Booking system for holistic health that lifted booking rate 15%. Real-time customer↔vet chat on Socket.io. Introduced a self-review practice that cut bugs 10%.
LetsDunch — Frontend Engineer · Feb 2020 – Dec 2023 Led 2 juniors building a mentorship collaboration platform. Rebuilt mentor search + time-slot booking, +8% conversion.
🎓 B.Sc. Computer Engineering — Port City International University
- 🧪 Eval harnesses — Langfuse, Braintrust, custom LLM-judge pipelines with regression gates
- 🕸 Multi-agent orchestration — supervisors, handoffs, parallel agents, durable execution
- 🎨 Generative UI — streaming React components straight out of tool calls
- 🔭 On-device inference — small models via WebGPU + Transformers.js
I post regularly on LinkedIn about shipping LLM features — retrieval that actually retrieves, agents that fail safely, and the eval work nobody puts in the demo. The parts that break, not the happy path.
If you're building with LLMs — RAG that has to be accurate, agents that have to be safe, or evals you don't have yet — I'm interested.
Anyone can get an LLM to answer. The job is knowing whether the answer was right.