Skip to content
View subhajeet-sapient's full-sized avatar

Block or report subhajeet-sapient

Block user

Prevent this user from interacting with your repositories and sending you notifications. Learn more about blocking users.

You must be logged in to block users.

Content in all repositories owned by your account will be closed.
Maximum 250 characters. Please don’t include any personal information such as legal names or email addresses. Markdown is supported. This note will only be visible to you.
Report abuse

Contact GitHub support about this user’s behavior. Learn more about reporting abuse.

Report abuse
subhajeet-sapient/README.md

Subhajeet Dey

SDM · Enterprise Agentic Framework · Publicis Sapient
Harness engineering · cost and performance optimization for LLM systems

Publicis Sapient Agentic Systems Inference Economics


Harness engineering

I work on the Enterprise Agentic Framework team, building the harness layer that sits between application code and the models underneath it — provider abstraction, execution orchestration, caching, failure handling, and the accounting that tells you what any of it actually cost.

The model gets the attention. The harness decides whether the system is usable: whether a run finishes in fifteen minutes or four hours, whether a provider outage degrades or takes the whole job down, whether a repeated request costs full price or nothing at all.

What that involves

Provider abstraction. One calling surface across gateways and vendors, with model aliasing, per-model capability differences, and graceful degradation when a provider rejects a parameter it doesn't support — the request should lose the hint, not fail.

Execution and concurrency. Turning serial pipelines into bounded parallel ones without losing determinism or result ordering. Most long-running inference jobs are latency-bound, not compute-bound, and spend the majority of their wall-clock time waiting.

Failure semantics. Timeouts on every network path, bounded retries with backoff, and — the one that bites hardest — never letting a failed call be recorded as a completed one. A resumable pipeline that caches its own errors will happily finish, fast and wrong.

Caching and resumption. Content-addressed reuse so interrupted work resumes instead of restarting, with cache keys that actually reflect every input that changes the output.

Contact

subhajeet.dey@publicissapient.com

Popular repositories Loading

  1. subhajeet-sapient subhajeet-sapient Public

  2. RouterArena RouterArena Public

    Forked from RouteWorks/RouterArena

    RouterArena: An open framework for evaluating LLM routers with standardized datasets, metrics, an automated framework, and a live leaderboard.

    Python