01 / LLM product
Sanchit Bhalla / AI engineer
I build AI that still works after the demo.
LLM features, RAG systems, and agents with real tools, taken from prototype to production by someone who also owns the backend, the pipeline, and the server they run on.
What I build
AI inside real products.
-
01
LLM product features
AI that lives inside the product, not beside it: streaming responses, structured output your code can trust, prompts under version control, and usage limits that protect your margins.
Prompt design / Structured output / Streaming / Metering
-
02
RAG and semantic search
Answers grounded in your own data. Chunking, embeddings, hybrid search, and reranking tuned against real questions, with citations so people can check the source.
Embeddings / Vector search / Hybrid retrieval / Reranking
-
03
Agents and tool calling
Agents that take real actions through well-defined tools, with validated inputs, bounded loops, and a person in the loop wherever a mistake would be expensive.
Tool calling / MCP / Orchestration / Human review
-
04
Running it in production
The work after the demo: evals that catch regressions, tracing for every run, caching and model routing for cost and latency, and fallbacks for when a provider goes down.
Evals / Tracing / Guardrails / Cost control
How I work
Every step is one you can inspect.
An example of how I structure an agent run, so you can see what it did, why, and where a person stays in control.
- 01 REQUEST Turn last week's client calls into a task list
- 02 PLAN Needs the call notes, then a write to the task system
- 03 RETRIEVE Hybrid search over notes, rerank, keep the best matches
- 04 CALL TOOL create_tasks(items, owner, due_date)
- 05 VALIDATE Check arguments against the schema, retry once on failure
- 06 REVIEW Pause for approval before anything is written
- 07 RESPOND Summary with citations back to each source note
- 08 LOG Save the trace, add the run to the eval set
Projects
What I've shipped.
Two AI products and a Rust networking tool, all live right now.
Experience
Where I've worked.
-
Jan. 2025 - Present
Fullstack Developer, Ebslon Infotech
Developed ERP systems end to end. Managed DevOps, CI/CD pipelines (Jenkins, GitHub Actions), and VPS infrastructure (Hetzner, Contabo). Implemented real-time features with WebSockets and Redis caching.
-
Nov. 2021 - Jan. 2022
Frontend Intern, Bestor
Designed and improved UI/UX of websites in Shopify. Collaborated with developers and SEO team to enhance user experience and site performance.
Stack
What I work with.
- Model APIs
- OpenAI, Anthropic, Gemini
- Orchestration
- Vercel AI SDK, LangChain, LangGraph
- Agents
- Tool calling, MCP, Structured output
- Retrieval
- Embeddings, pgvector, Pinecone, Hybrid search, Reranking
- Quality
- Evals, Tracing, Guardrails, Prompt caching
- Languages
- TypeScript, Python, Rust, Go
- Data and real-time
- PostgreSQL, Redis, WebSockets
- Infrastructure
- GitHub Actions, Jenkins, Hetzner, Contabo
About
A model is the easy part. I care about everything around it: the data it sees, the tools it can touch, how you find out when it gets worse, and an interface people actually trust. Coming from full-stack and infrastructure work, I build all of that myself.
Contact
Have an AI feature in mind?
If you have a product, a role, or a hard problem, tell me about it.