Waqas Sharif

AI engineer in Karachi, UTC+5, working on production LLM agent systems and the evaluation harnesses that keep them reliable.

WHAT I BUILD

Primary architect of Rahzaan, a deployed agentic career guidance platform for secondary school students. It runs a 24-node parallel LangGraph pipeline with 52 typed state fields and concurrent-write reducers, cutting first-response latency from 20-25 seconds to 3.7. The backend is 22 REST endpoints in FastAPI with SSE streaming, checkpoint persistence to PostgreSQL, and per-node model routing so any node's LLM is swappable through environment config alone.

RELIABILITY IS THE PART I CARE ABOUT

I built an evaluation harness of five named personas and a ten-scenario matrix producing 60 per-turn traces, plus 19 behavioural probes using paired before-and-after measurement and single-criterion isolation, behind a 46-file suite of 592 automated tests.

That harness caught something I had wrong. I assumed the system over-asserted because of thin data. Measuring across five completeness levels showed the opposite: the thinnest profile produced the most honest output, declining outright rather than fabricating. The real trigger was internal signal divergence, the model inventing a bridge to reconcile conflicting inputs. The guard I had planned would have fixed nothing.

RESEARCH

A behavioural study of production LLM workflows naming three reproducible failure modes, and a measured ~30% retrieval error rate when context follows rather than precedes the instruction.

HOW I WORK

AI-natively. An architecture conversation produces versioned specifications, an implementer works from them under a written rules file, and traces come back to close the loop. 334 session records across four months, with the model and reasoning tier attributed per session.

BE Software Engineering, KIET Karachi, 2026. CGPA 3.47/4.0.