Backend & AI-systems engineer · Ahmedabad, India
I build systems where software does real, verifiable work — and I care most about the unglamorous half: hybrid retrieval that actually ranks, isolated execution, and state that survives a restart.
🔨 Forge — an AI software engineering platform
Not a coding assistant — a platform that owns the loop from task to verified change: understand the codebase → plan → implement in isolation → validate → debug → review → remember the outcome as project memory.
A modular monolith plus an isolated execution worker: the control plane (API + orchestrator + memory) is deliberately separated from the execution plane (worker + sandbox), so untrusted generated code never runs on the API host.
FastAPI · Next.js · PostgreSQL + pgvector
🔍 Problem Finder — hybrid search over coding problems
A hybrid search engine that merges three retrieval strategies — semantic (all-MiniLM-L6-v2, 384-d), full-text, and trigram fuzzy — using Reciprocal Rank Fusion, so a half-remembered problem title still surfaces the right result. Includes Norvig spell-correction for "did you mean".
PostgreSQL + pgvector (FTS · trigram · HNSW indexes) · FastAPI · React + Vite — architecture, data-pipeline, API and deployment docs in-repo, with a pytest suite.
- Retrieval is a ranking problem, not a lookup problem. Precision comes from fusing signals, not from picking one clever one.
- Verification over vibes. A change isn't done when tests go green — it's done when I can explain why it's correct.
- Design before code. Architecture and data-flow docs get written first; the diff follows.
- Isolation as a default. Untrusted code and untrusted input get their own plane.
Open to backend / AI-systems roles — prakashsankhla0601@gmail.com