Research Engineer | LLM Post-Training, Evaluation & Reliable Agent Systems

I work at the intersection of reward modeling, reasoning-model evaluation, verifiable post-training, and agentic AI engineering. My projects emphasize reproducible experiments, trustworthy feedback signals, controlled baselines, and failure analysis—not just headline benchmark scores.

• Stable Outcome Reward Modeling via Pairwise Preference Learning — Studied pointwise reward-model instability and pairwise outcome supervision. Reported 96.3% held-out pairwise accuracy (90% bootstrap CI: 95.3–97.1%) within the project's evaluation setup. Preprint · Model · Dataset
• Verifiable Agentic RL Lab — Built a strict symbolic verifier and evaluated a small reasoning model using sampling, LoRA, and GRPO. On 200 frozen Countdown tasks, untouched-model greedy success was 4.5%, versus 23.0% best-of-8 verifier selection (an inference-time comparison). Controlled RLVR runs did not establish reliable held-out correctness gains; the negative result and limitations are documented.
• MAGI-RM — Experimental shared-encoder outcome/process reward-model pipeline with training utilities, labeled reasoning-chain data, and quality-control tooling. Trained-model performance is not yet established in the public repository.

• Nexus RAG — Domain-specific retrieval-augmented question answering with dense + BM25 retrieval, reranking, citations, refusal handling, and evaluation workflows.
• MARCS — Research prototype for LLM-assisted code review, journaled patch application, recovery, safety checks, and traceable reports; not presented as production-hardened.
• SERA — Research-paper ingestion pipeline with arXiv retrieval, PDF processing, embeddings, and metadata storage.

Current direction: execution-verified Python code repair and agent training, starting with authentic-task feasibility and independent verification; supervised adaptation and RL are conditional on evidence.

Technical focus: Python · PyTorch · Transformers / PEFT / TRL · LoRA / GRPO · LLM evaluation · RAG / retrieval · FastAPI · Docker

Interested in: research engineering and applied AI engineering roles spanning LLM evaluation, alignment/post-training, and reliable agents.

LinkedIn · Hugging Face · Email: akleshmishra7@gmail.com