sahildahiya.me / home
Most recently at Workleap, previously five years at Microsoft. Most of what I've shipped is on the eval side: frameworks, A/B tests on model choices, traces that survive a postmortem.
I run two projects of my own to keep the work close to the runtime.
At Workleap I worked on AI assistants and the systems that made them measurable: routing, RAG pipelines, eval frameworks, A/B testing on model choices, and conversation anonymization for product analysis. I also owned a feature end to end: a natural-language query engine that lets IT admins ask governance questions in plain English, analyzes results, and suggests follow-ups they didn't think to ask. Built MCP integrations that extended the assistant across three product lines.
Before that, five years at Microsoft on anomaly detection, forecasting, experimentation, and device health systems across Windows and data center ops. The work I'm proudest of is an anomaly detection system that caught arcing precursors in electrical telemetry across the global data center fleet before they became incidents.
01 / active system
A recommendation system built on RAG for philosophy reading. Each conversational turn retrieves passages, narrows to evidence, generates a cited answer, and updates a reading list so recommendations improve across turns. Search, passage-level conversation, essay drafting, all in one workflow.
02 / active runtime
A terminal coding agent built around a Python runtime, JSON-over-stdio RPC, and a first-party Go interface. The project is focused on keeping the backend contract explicit, the TUI thin, and the runtime strict enough to support real coding sessions. I replaced subprocess-per-call execution with a long-lived Go worker (1,300x warm-read speedup) and run evals across multiple model cohorts: 47.4% on Terminal-Bench 2 with GLM-5 (public submission), higher on GPT-5.4.
Short notes I keep so model choices, eval runs, and deployment decisions stay findable later.
the core problem is not one-shot answer generation; it is recommending the next passage worth reading, with retrieval and synthesis used to keep that recommendation grounded
the session was not stuck because `git` was slow; it was stuck because the backend was pushing too many large updates through a small Windows pipe
it is not; an agent needs both conversation memory and a separate runtime-framing baseline for the next run
between runs, compaction decides what past survives, but a long-running coding agent also needs a separate runtime-framing baseline so the next run starts under the right conditions
Want to talk about AI evaluation, agent runtimes, or anything above? dahiya.sahil.89@gmail.com