WORK / AGENTIC ENGINEERING ASSISTANT
Agentic Engineering Assistant
A production RAG and tool-calling platform: a three-pass request architecture that keeps prompt size flat as tools are added, with correctness enforced by the backend instead of the prompt.
ROLE
Backend & AI engineer, Design Intelligence LLP
STACK
FastAPI · OpenAI function calling · Pinecone · Pydantic
TIMELINE
2025 - ongoing
STATUS
In production
THE PROBLEM
An engineering assistant has to call the right tool for calculations where a wrong answer matters. The naive design attaches every tool schema to every request: prompts grow with each tool you add, and tool selection gets worse as the list gets longer.
The harder problem is trust. A model told to confirm inputs before a safety-critical calculation will usually do it - and occasionally won't. "Usually" isn't acceptable there. This is described at the architecture level only; client and domain details are under NDA.
CONSTRAINTS
- -Safety-critical calculations need explicit user confirmation, every time
- -The tool catalogue keeps growing - cost and accuracy can't degrade with it
- -Multi-turn configuration flows, but the service has to stay stateless
- -Re-ingesting a document must replace its chunks, never duplicate them
ARCHITECTURE
Each request runs in three passes. A routing call with no tool schemas attached picks one active tool; execution sees only that tool's schemas; a dedicated zero-temperature call formats the result. Deterministic gates in the backend sit between the model and anything it wants to run.
- 01Route - one call with no tool schemas resolves a single active tool
- 02Execute - only that tool's schemas are sent, so prompt size doesn't depend on tool count
- 03Gate - backend checks re-derive tool disambiguation and user confirmation from the transcript
- 04Format - a dedicated temperature-0 call shapes the final answer
- 05State - recovered from markers and embedded JSON in the transcript; no session store, no database
- 06RAG - Pinecone, text-embedding-3-small (1536-d), 700-word chunks with 120-word overlap
- 07Rerank - blended score, 0.7 vector similarity + 0.3 lexical overlap
DECISIONS
1.Route before you execute
Splitting tool selection from tool execution means the expensive, schema-heavy call only ever carries one tool. Adding the twentieth tool costs the router a line of description, not every request a full schema.
2.Enforce correctness in the backend, not the prompt
The backend re-derives from the conversation whether the user actually confirmed the inputs. A model that skips the confirmation ritual on a safety-critical calculation is rejected rather than trusted - the prompt asks for good behaviour, the code guarantees it.
3.Stateless, with state recovered from the transcript
Multi-turn configuration flows rebuild their in-progress state by parsing markers and embedded JSON from the transcript itself. No session store means nothing to expire, migrate or keep in sync across instances.
4.Retrieval as a tool, not a pre-step
RAG is exposed as a tool the model invokes when it needs documents, rather than an unconditional lookup before every call. Deterministic document IDs make re-ingestion idempotent.
RESULTS
3-pass
Route · execute · format
0
Session stores or databases
0.7 / 0.3
Vector / lexical rerank blend
what I'd do differently…
I'd build a labelled set of real requests for the router in the first week. Routing is the one call every request depends on, and without an evaluation set every change to its prompt is a guess.
quick answers
Quick answers about Agentic Engineering Assistant
- What is Agentic Engineering Assistant?
- A production RAG and tool-calling platform: a three-pass request architecture that keeps prompt size flat as tools are added, with correctness enforced by the backend instead of the prompt.
- What is Agentic Engineering Assistant built with?
- Agentic Engineering Assistant is built with FastAPI, OpenAI function calling, Pinecone, Pydantic.
- What was SK Rohan Parveag's role on Agentic Engineering Assistant?
- Backend & AI engineer, Design Intelligence LLP.
- What is the current status of Agentic Engineering Assistant?
- In production.