Open to senior AI engineering roles

I build AI systems that hold up in production: retrieval, tool-calling agents, guardrails, and the maps that make them legible.

Akshay Prabhakant Founding Product Engineer, RunWhen Pune, India

Founding product engineer at RunWhen, building production retrieval, tool-calling agents, and the guardrails and evaluation that keep them honest. Before that, four years building ML systems in logistics — including an ETA model that moved on-time shipments from 83% to 92%.

All of it is the same problem in different clothes — you cannot fix a system you cannot see — which is why half of what I make ends up as a navigable map of something too large to hold in your head.

Currently
Production RAG over a Neo4j vector store — hybrid dense + BM25 fused with reciprocal rank fusion — a single tool-calling agent on Google ADK, and multi-provider LLM routing through a LiteLLM proxy
Research
Three papers in computational biophysics, 14 citations
Education
IIIT Hyderabad — B.Tech Computer Science; MS by Research, Computational Sciences
Toolkit
Python · Google ADK · LiteLLM · RAG and hybrid retrieval · Neo4j · Qdrant · Presidio guardrails · FastAPI · Kubernetes · Docker · GCP
01

Work

02

Research

Google Scholar →
14Citations
2h-index
3Papers
  1. Nucleic Acids Research Q1 IF 15.0

    Mapping the recognition pathway of cyclobutane pyrimidine dimer in DNA by Rad4/XPC

    N Jakhar, A Prabhakant, M Krishnan 51(19) 10132–10146 · 2023 3 citations DOI ↗ PubMed ↗

    Rad4/XPC is a damage sensing protein that recognizes and repairs CPD lesions with high fidelity. However, the molecular mechanism of how Rad4/XPC interrogates CPD lesions remains elusive. Emerging viewpoints indicate that the association of Rad4/XPC with DNA, the insertion of a lesion-sensing β-hairpin into the lesion site and the flipping of CPD's partner bases are essential for damage recognition.

  2. Journal of Chemical Information and Modeling Q1 IF 6.4

    Response of Terahertz Protein Vibrations to Ligand Binding: Calmodulin–Peptide Complexes as a Case Study

    B Varvdekar, A Prabhakant, M Krishnan 62(7) 1669–1679 · 2022 9 citations DOI ↗ PubMed ↗

    Ligand binding alters the nature and distribution of collective terahertz protein vibrations. The ligand-induced changes in these vibrations contribute to the binding entropy and to the overall thermodynamic stability of the resultant protein–ligand complexes. The study examines the low-frequency response of calcium-loaded calmodulin to five different ligands, with and without water, using normal-mode analysis and molecular dynamics.

  3. Journal of Chemical Information and Modeling Q1 IF 6.4

    Allosteric Response of DNA Recognition Helices of Catabolite Activator Protein to cAMP and DNA Binding

    A Prabhakant, A Panigrahi, M Krishnan 60(12) 6366–6376 · 2020 2 citations DOI ↗ PubMed ↗

    Binding of cAMP to CAP triggers allosteric communication between its cAMP binding domains and DNA binding domains. That communication entails repositioning the DNA recognition helices in the DNA binding domain to dock favourably to the target DNA. The study uses molecular dynamics and umbrella sampling to map the allosteric pathways involved.

03

Experience

  1. RunWhen

    Jul 2025 — Present

    Pune, India

    RAGGoogle ADK LiteLLMNeo4j Presidiohybrid search

    Founding Product Engineer

    • Built a production RAG retrieval service over a Neo4j vector store — OpenAI text-embedding-3-small embeddings, hybrid dense + BM25 retrieval fused with reciprocal rank fusion, and sentence-level chunking.
    • Architected a conversational agent on Google ADK that exposes the workspace as a virtual filesystem tool interface (ls, cat, grep, search over a Neo4j-backed graph), replacing a prior 8-sub-agent, 74-tool design with a single tool-calling agent.
    • Built multi-provider LLM orchestration through a LiteLLM proxy spanning Gemini, GPT-4.1 and OpenAI, with workspace-scoped model routing, Vault-backed credentials and per-turn token accounting.
    • Built the guardrails and the evaluation around them: a fail-closed PII and secret scrubber, and a custom ADK agent-eval harness with rubric-based LLM-judge scoring and a user simulator for multi-turn regression checks.

  2. XPO Logistics

    Jul 2021 — Mar 2025

    Pune, India

    GeminiRAG document extractionrandom forests A/B testingBigQuery

    Data Scientist 2 · Jul 2022 – Mar 2025

    • Built a Gemini-flash RAG chatbot over XPO's external help centre, then extended the same retrieval approach to pricing-agreement PDFs with heterogeneous layouts at 88% accuracy.
    • Built an end-to-end shipment ETA pipeline on random-forest regressors: 90% day-to-delivery accuracy against 84% before, lifting on-time shipments from 83–87% to 88–92%.
    • Validated the rollout the hard way — a matched 12-site UAT cohort across four states on hourly throughput, trailers and workforce, then a two-proportion z-test before full deployment.
    • Quantified price elasticity and a willingness-to-pay index in BigQuery and DataStudio, driving a 9% volume increase and 7% revenue lift.

    Data Scientist · Jul 2021 – Jun 2022

    • Traced the causes of damaged shipments through BigQuery — weight, handling counts, distance, NMFC class, workforce type, in-house versus third-party transport — contributing to a 19% reduction.

View the full résumé ↗

04

Elsewhere

Open to conversations about agent infrastructure, evaluation and data visualisation — and to anyone who wants to argue about the election atlas.

Akshay Prabhakant · Pune, India