Forward Deployed AI Engineer · Agentic Systems, GraphRAG & Applied ML

RushikeshRushikesh GholapGholap

Architecting durable multi-agent control planes, two-plane 3-hop GraphRAG systems, capability-scoped MCP tool servers, and low-latency GPU inference in regulated enterprise environments.

< 2 min
Clinical Dossier SLA
BioNTech · 4-Agent Consensus (1.60s online)
96.4%
3-Hop GraphRAG Recall
Spark GraphFrames vs. 54.2% Flat Vector
0.68s
P95 Voice Turnaround
Post-VAD · GPU Whisper A10 · 5x Capacity
500K+
Daily Vector & Risk Txns
Vertex AI · EKS/GKE · 30+ Prod Models
BioNTech · Multi-Agent Control Plane

Autonomous Agents18.4-minute synthesis in under 2 minutes

Spark GraphFrames · Mosaic AI

Multi-Hop GraphRAG96.4% 3-hop recall vs. 54.2% flat vector

Forward deployed executionFrom ambiguous customer schemas to governed production AI systems

2019–2026 track record
01PHASE 01 · EMBEDDED TECHNICAL ARCHAEOLOGY

Ambiguous Customer Workflows → Typed Ontologies & Evals

Embed directly with clinical, banking, and operations leads to map undocumented schemas, regulatory PDFs, and legacy APIs into permission-aware entity graphs and golden evaluation benchmarks before writing agent loops.

BioNTech Oncology ProtocolsPDB Bank 105-Signal KYCAlmavest 200+ Lenders
02PHASE 02 · DURABLE CONTROL PLANES & MCP

Fault-Tolerant State Machines, MCP Sidecars & Log-Odds Consensus

Separate deterministic workflow durability (Temporal, checkpointed LangGraph, OAuth 2.1 capability-scoped MCP tool servers) from calibrated multi-agent reasoning and adversarial NLI entailment gates.

99.1% Citation PrecisionIdempotent MCP Tool Sagas3-Hop GraphFrames Retrieval
03PHASE 03 · VPC SERVING & UNIT ECONOMICS

Low-Latency GPU Inference, CI Release Gates & Day-2 Telemetry

Deploy async FastAPI services, quantized vLLM with prefix KV caching, LiteLLM multi-provider routing, and OpenTelemetry / Promptfoo CI release gates across regulated Databricks, AWS, and GCP environments.

-85% P99 Inference Latency0.68s P95 SIP Voice Agent< 10 min Loan Underwriting

Interactive architectureTwo-plane GraphRAG, durable MCP execution, and calibrated consensus

INTERACTIVE WORKBENCH · 01

Multi-Hop GraphRAG & Durable Control Plane vs. Naive Baselines

QUERY

Identify cross-trial dosing modifications when mRNA neoantigen therapy exhibits Grade ≥2 cytokine release in HLA-A*02:01 cohorts.

Public Literature Note: Illustrative walk-through constructed strictly from public ClinicalTrials.gov (NCT04526899) and NCI CTCAE v5.0 literature; zero proprietary sponsor data.
RELIABILITY / RECALL96.4%
EXECUTION TOPOLOGY3-Hop GraphFrames + Mosaic Cache
SYNTHESIS TRACEOffline Spark GraphFrames materializes 3-hop paths into Delta Lake; online Mosaic AI cache resolves HLA-A*02:01 → Amendment 4B → step-down prime/boost with 99.1% citation precision.
BIONTECH · ORCHESTRATION

Calibrated Multi-Agent Consensus

18.4m → <2m

Parallel 4-agent ensemble with Platt-calibrated log-odds pooling S(y) = Σ wᵢ · log(pᵢ / (1 - pᵢ)) and NLI contradiction pruning.

GraphRetriever-αw=0.34 · 97%
14 grounded entity-relation triplets verified
BioClinical-βw=0.29 · 95%
Cohort inclusion & adverse event delta aligned
AdversarialCritic-γw=0.22 · 98%
0 unsupported claims · 2 redundant citations pruned
ConsensusArbiter-Ωw=0.15 · 96%
Online consensus in Round 2 (1.60s · <2m full dossier)
VLLM · NVIDIA A10 · KUBERNETES LLMOPS

Low-Latency VPC Serving & Telemetry

Quantized GPU inference, full-duplex SIP voice AI, and HNSW vector pipelines governed across regulated enterprise cloud pods.

-85%
P99 Inference Latency & Token CostQuantized vLLM (AWQ/FP8) + prefix KV caching
0.68s
P95 Post-VAD Voice TurnaroundCTranslate2 Whisper on NVIDIA A10 + Asterisk SIP
500K+
Daily Vector & Warehouse TxnsHNSW + BigQuery/Redshift · 30+ governed models

Selected system specificationsLakehouse GraphRAG, durable MCP runtimes, GPU voice AI, and explainable ML

Explore all 12 system specs
01 · BIONTECH · HEALTHCARE GENAI
ENTERPRISE PRODUCTION · BIONTECH

Enterprise Healthcare GenAI Platform — Two-Plane GraphRAG & Consensus

Architected a production Healthcare Generative AI platform on Databricks for BioNTech—separating offline Spark GraphFrames 3-hop entity-relation indexing from online Mosaic AI retrieval and 4-agent calibrated log-odds consensus to compress clinical protocol synthesis from weeks to under 2 minutes at 99.1% citation precision.

< 2 min
Dossier Synthesis
Down from multi-week manual review (1.60s online)
99.1%
Citation Precision
Platt-calibrated 4-agent log-odds consensus
96.4%
3-Hop Recall
Spark GraphFrames vs. 54.2% flat cosine
ARCHITECTURE PIPELINEForward Deployed AI Engineer
01Layout PDF Ingestion
022-Plane GraphRAG
034-Agent Log-Odds
04vLLM & HITL Audit
02 · AGENTIC INFRA · MCP & TEMPORAL
2026 REFERENCE ARCHITECTURE

MCP & Durable Agentic Control Plane — Fault-Tolerant Enterprise Runtime

A 2026 forward-deployed reference architecture separating deterministic workflow durability (Temporal.io + checkpointed LangGraph) from capability-scoped Model Context Protocol (MCP) tool sidecars, multi-provider LiteLLM routing, and CI/CD regression eval gates.

0
Duplicate Mutations
SHA-256 idempotency ledger across replays
< 35ms
MCP Policy Overhead
OAuth 2.1 On-Behalf-Of capability check
98.2%
CI Eval Pass Gate
Promptfoo + OTel trajectory assertions
ARCHITECTURE PIPELINESystems & Applied AI Architect
01OIDC & RBAC Gateway
02Temporal + LangGraph
03Sandboxed MCP Tools
04LiteLLM & Eval Gate
03 · MARKYTICS · VOICE AI & TELEPHONY
ENTERPRISE PRODUCTION · MARKYTICS

Real-Time Multilingual Voice Agent — GPU Whisper ASR & SIP Telephony

Architected and deployed a production multilingual contact-center voice agent at Markytics—combining Asterisk PBX duplex RTP streaming, Silero VAD endpointing, INT8/FP16 CTranslate2 Whisper ASR on NVIDIA A10 GPUs, and streaming Google Neural2 SSML TTS to achieve 0.68s P95 post-VAD turnaround and 5x concurrent call capacity.

0.68s
P95 Turnaround
Post-VAD endpoint to first outbound RTP packet
5x
Call Capacity
Batched INT8/FP16 Whisper on NVIDIA A10 GPUs
-67%
Resolution Time
Automated multilingual dialogue across 4 languages
ARCHITECTURE PIPELINELead Data Scientist (Technical Lead)
01Asterisk SIP & VAD
02A10 CTranslate2 ASR
03Intent & Slot FSM
04Streaming SSML TTS
04 · OPEN-SOURCE · GRAPHRAG & CRAG
OPEN-SOURCE WORKBENCH

FactChain — Corrective GraphRAG & Multi-Hop Claim Verification

An open-source Corrective GraphRAG and multi-agent claim verification workbench uniting LightRAG, Neo4j, LlamaIndex, PostgreSQL (pgvector + BM25 RRF), and programmatic DSPy verifiers to decompose complex claims and enforce NLI source entailment.

94.8%
3-Hop Recall
Entity-graph + RRF vs. 48.0% flat cosine
0.89s
Cached Traversal
Neo4j / LightRAG neighborhood lookup
DSPy + NLI
Self-Correction
Query rewrite when entailment < 0.85
ARCHITECTURE PIPELINEAuthor & Lead AI Engineer
01Entity Graph Indexer
02Neo4j + Hybrid RRF
03DSPy Claim Synthesizer
04NLI Critic & Retry
FORWARD DEPLOYED AI ENGINEERING · AVAILABLE FOR L4 / L5 IMPACT

Ready to ship fault-tolerant AI from whiteboard to production?

Architecting durable multi-agent systems, Spark GraphFrames retrieval, OAuth-scoped MCP integrations, and low-latency GPU inference for regulated enterprise teams.

RGX ARCHITECTURE CONCIERGELOCAL SPEC INDEX
Search across 12 system specifications, BioNTech 3-hop GraphRAG, Temporal + MCP durable execution, 0.68s GPU Whisper voice AI, or 2019–2026 experience.