Skip to content
View rupeshpoojary9's full-sized avatar
1 follower · 2 following

Achievements

Achievement: Pull SharkAchievement: QuickdrawAchievement: Pair ExtraordinaireAchievement: YOLO

Achievements

Achievement: Pull SharkAchievement: QuickdrawAchievement: Pair ExtraordinaireAchievement: YOLO

Block or report rupeshpoojary9

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
rupeshpoojary9/README.md
Rupesh Poojary — AI Engineer

LinkedIn   X   DEV.to   Medium   Email   GitHub   views

 

 Selected work

I build the layer that makes language models trustworthy: orchestration that verifies itself, retrieval you can measure, and guardrails that hold. Every project ships with a real eval harness, because "it felt better" is not a metric.

▸ multi-agent-orchestrator

A meta-agent plans a task DAG, routes steps to web and typed-API agents, runs a plan-verify-iterate loop, and synthesizes new sub-agents at runtime.

100% · 26-task eval   100% verifier catch

py stars

▸ rag-eval-benchmark

BM25 vs dense vs hybrid (RRF) vs hybrid + rerank over a hand-labelled corpus. Local models, no API key required.

hybrid recall@5 0.925   BM25 0.738

py stars

▸ bulwark

Local-first LLM guardrails: layered PII, injection, secrets, and schema validators with a built-in eval harness.

stops unsafe I/O before it reaches the model

py stars

▸ genai-refinement-advisor

Recommends the cheapest refinement level (prompt vs RAG vs agent vs fine-tune) that clears a quality bar, via cost-performance Pareto benchmarking.

data-backed "which approach do we need"

py stars

▸ poorjev

Open-source, local "System One" decision layer for LLM apps: typed decisions with provably calibrated confidence. No API key, no waitlist.

ECE 0.170 → 0.071   zero accuracy loss

py stars

▸ opie

Image to validated, personalized food-product intelligence: multi-agent extraction, a misleading front-of-pack claim detector, and a data flywheel that lifts attribute F1 across rounds.

misleading recall 0.936   F1 0.966

py stars

 

 Stack

 

 Activity

contribution snake animation

Based in Mumbai, India  ·  open to conversations about hard LLM reliability problems.

Pinned Loading

  1. awesome-llm-evals awesome-llm-evals Public

    A curated list of frameworks, benchmarks, and research for evaluating LLMs and LLM-powered systems.

  2. poorjev poorjev Public

    Open-source, local Jev alternative: a System One decision layer with provably calibrated confidence (ECE 0.170→0.071). Typed decisions, runs offline, no API key, no waitlist.

    Python 10

  3. bulwark bulwark Public

    Local-first LLM guardrails toolkit: layered PII/injection/secrets/schema validators with an eval harness.

    Python

  4. genai-refinement-advisor genai-refinement-advisor Public

    Decision-support tool that recommends the cheapest GenAI refinement level (prompt/RAG/agent/fine-tune) that clears a use case's quality bar, via cost-performance Pareto benchmarking of candidate fo…

    Python

  5. multi-agent-orchestrator multi-agent-orchestrator Public

    Self-extending multi-agent orchestrator: a meta-agent plans a task DAG, routes steps to a web agent and a typed API agent, runs a plan-verify-iterate loop, and synthesizes new sub-agents at runtime…

    Python

  6. rag-eval-benchmark rag-eval-benchmark Public

    Reproducible RAG retrieval benchmark: BM25 vs dense vs hybrid(RRF) vs hybrid+rerank over a hand-labelled corpus. Reports recall@k, MRR, nDCG + answer faithfulness. Local models, no API key.

    Python