Retrieval
Problem
Weak grounding creates confident but incorrect responses.
Decision
Hybrid retrieval + contextual memory + evaluation datasets.
Tradeoff
Higher infra complexity for better answer trust.
Engineering
The technology I trust, test, and question.
I write and build from the engineering side of AI: stateful agents, retrieval quality, backend contracts, model runtime constraints, deployment topology, and the tradeoffs that show up after a demo becomes a real system.
engineering.log
Primary Lens
Production AI systems
Current Focus
Agents, RAG, backend reliability
Architecture Style
Explicit state, typed APIs, observable flows
Audience
Engineers, founders, platform teams
This page is for engineers who want to inspect my stack choices, developers who want a practical signal, and clients who need to understand how I make architecture decisions.
Architecture Radar
My current technology radar for production AI.
I use this as an engineering ledger: what I trust in production, what I am testing in labs, what I am studying carefully, and what I avoid when reliability matters.
Orchestration
Agent state, tool routing, memory boundaries, retries, and human approval gates.
Models
Cloud and local runtimes evaluated against latency, privacy, cost, and reasoning depth.
Data
Retrieval quality, hybrid search, pgvector patterns, reranking, and grounding loops.
Platform
FastAPI services, container paths, observability, deployment handoff, and frontend surfaces.
100%
Drag to Pan | Scroll to Zoom
Interactive Legend (Hover to Spotlight | Click to Filter)
Adopt
24
Trial
7
Assess
2
Hold
1
Systems Directory
Adopt Ring (24)
Trial Ring (7)
Assess Ring (2)
Hold Ring (1)
ADR Technical Document
AI OrchestrationAdopt (Production-Ready)
LangGraph
> System Insight:
Manoj's primary choice for building stateful multi-agent systems. The checkpointer persistence pattern is vital for banking and business workflows.
[PROBLEM / CONTEXT]
Single-prompt loops fail on multi-step reasoning, resulting in brittle tool executions.
[ARCHITECTURAL DECISION]
Deploy LangGraph to model agent transitions as deterministic state machines with persistence layers.
[ENGINEERING TRADEOFFS]
Requires explicit state modeling in exchange for predictable, resumable execution paths.
Decision Ledger
The stack is only useful when the tradeoffs are visible.
I do not pick tools because they are fashionable. I start from failure modes: weak grounding, opaque agent loops, prototype backends, missing traces, slow feedback cycles, and platform handoff risk.
Decision map
How production constraints shape architecture choices
Orchestration
Problem
Single-agent flows fail on multi-step business operations.
Decision
LangGraph-style stateful agent routing with tool gating.
Tradeoff
More state modeling in exchange for reliability.
Backend
Problem
Prototype notebooks cannot support production workloads.
Decision
FastAPI services, async workers, queue-backed execution.
Tradeoff
More platform discipline for delivery velocity.
Observability
Problem
Teams cannot explain failures or regression causes.
Decision
Trace IDs, eval loops, latency/cost/error instrumentation.
Tradeoff
Instrumentation overhead for operational confidence.
System Maps
Where the radar turns into deployable systems.
These are the recurring patterns I come back to when an AI product has to leave a notebook and become a service that developers can own, debug, and improve.
Multi-Agent Systems
LangGraph-based workflows with explicit state, tool routing, memory, fallbacks, and evaluation.
- 01
Intent
- 02
Planner
- 03
Agent State
- 04
Tools
- 05
Human Gate
- 06
Trace
Tradeoffs
state visibilitytool safetyretry behaviorhuman control
RAG Reliability
Retrieval pipelines designed for grounding quality, latency budgets, observability, and regression testing.
- 01
Corpus
- 02
Chunking
- 03
pgvector
- 04
Hybrid Search
- 05
Rerank
- 06
Evals
Tradeoffs
chunkingrankinggroundinglatency
FastAPI AI Backends
Async Python services, model gateways, queues, trace IDs, and deployment paths for AI product teams.
- 01
API
- 02
Queue
- 03
Workers
- 04
Model Gateway
- 05
Store
- 06
Observability
Tradeoffs
async workloadsAPI contractscost controlsdeployment
Stack
Hands-on across the layers where AI systems break.
My work sits across orchestration, retrieval, model runtime, backend services, infrastructure, and frontend delivery. The breadth is intentional because production AI rarely fails in only one layer.
Agent Runtime
LangGraphLangChainGoogle ADKMCP / ACP / UCP
Retrieval
pgvectorPostgreSQLMongoDBHybrid SearchPinecone
AI Backend
PythonFastAPIAsync workersModel gatewaysTyped APIs
Model Layer
OpenAIClaudeVertex AIAWS BedrockOllamavLLM
Platform
DockerKubernetesOpenShiftNVIDIA Run:AIAWS / GCP
Frontend Systems
Next.jsReactTypeScriptMicrofrontendsModule Federation
Trajectory
My engineering path started before the AI layer.
The AI work is built on years of frontend architecture, enterprise platform delivery, microfrontends, backend APIs, regulated banking systems, and cloud-native engineering.
2023 - PRESENT
AI Systems Architect
Publicis Sapient
Leading the architecture design and POC-to-production lifecycle of enterprise Generative AI platforms, multi-agent orchestrations, and high-recall retrieval systems.
CORE ARCHITECTURAL FOCUS
Designing resilient state-machine graphs (LangGraph), secure Cloudflare egress proxy tunnels, pgvector database configurations, and context compaction pipelines that survive enterprise audits.
PRODUCTION MILESTONES & AWARDS
- FS West Supernova Award for enterprise AI transformation and execution strategy
- Publicis Sapient Spot Award for GenAI platform innovation and team mentoring
- Reduced manual workflow lifecycle overhead by 40%+ using stateful agentic automations
- Accelerated POC-to-production lifecycle by nearly 2x across enterprise teams
Technology Radar
System.Scan_active
Orchestration & Logic
LangGraphLangChainGoogle ADKMCP / ACP / UCP schemas
Compute & Models
OpenAI (GPT-4o)Claude (3.5 Sonnet)AWS BedrockGCP Vertex AIOllamavLLMDeepSeekLLaMA
Databases & Retrieval
pgvectorPostgreSQLQdrantMongoDBHybrid Search
Infrastructure & Containers
DockerKubernetesRed Hat OpenShiftOpenShift AINVIDIA Run:AIGCP ComputeGCSCloudflare WAF
Package & Tooling
UVpnpmNPMGitHub ActionsCI/CD
VERIFIED PRODUCTION STACK
Manoj.Mukherjee_v1.0Publications
Research notes, papers, and engineering writing.
I treat writing as part of the engineering loop. It forces architecture decisions to become clear enough for other developers, technical buyers, and platform teams to evaluate.
Multi-Agent Systems2026
Agent-to-Agent Communication Systems
Medium / Research BriefArchitectural guidelines on multi-agent messaging protocols, state serialization, message schemas, and execution synchronization within LangGraph and Model Context Protocol architectures.
Architectural Impact:Modeled inter-agent execution loops and structured communication gateways to secure workspace routing.
LangGraphMCPState SynchronizationJSON-RPC
RAG Infrastructure2026
Advanced Retrieval-Augmented Generation (RAG)
Medium / Systems JournalAn in-depth systems analysis of hybrid retrieval execution. Focuses on combining dense vector search (pgvector/Qdrant) and sparse document keyword indexes (MongoDB), optimized through cross-encoder reranking.
Architectural Impact:Designed token-compaction formulas and hybrid search scoring formulas that reduced context payload sizes by 45%.
pgvectorQdrantCross-EncoderHybrid Search
Local Inference Platform2026
Real-Time AI Agents with LangChain & LangGraph
Research Hub / BlogResearch on low-latency local agent runtimes. Analyzes hardware-accelerated local execution constraints on Apple Silicon environments using Ollama, MLX, and vLLM gateways.
Architectural Impact:Built open-source benchmarks comparing time-to-first-token (TTFT) performance across unified memory lanes.
OllamaMLXvLLMLocal Hardware
Vision AI & Ingestion2025
AI + OCR + Vision Systems
Systems Engineering DigestA comprehensive design pattern for enterprise document ingestion. Combines layout-aware OCR parsers, layout bounding boxes, and multimodal LLMs to automate unstructured record extractions.
Architectural Impact:Architected a zero-loss layout parsing pipeline that coordinates structured entity checks with strict data boundaries.
Vision AILayout-ParserOCR IngestionMultimodal
Financial Systems Research2026
AI-Powered Stock Analysis using LangGraph & DeepSeek
Google Scholar / PublicationsSystems research analyzing automated financial market research. Features multi-agent routers executing complex technical audits, sentiment parsing, and algorithmic risk evaluations.
Architectural Impact:Modeled the financial risk checking rules node as a human-in-the-loop validation interrupt.
DeepSeekRedux SagaFinancial ChartsRisk Gating
Operating Philosophy
Read, explore, POC, repeat.
My engineering style is practical: study the system, build the smallest credible proof, measure the failure modes, and only then decide whether a tool deserves production trust.
Engineering Creed
Manoj Mukherjee
By heart a Software Engineer — Look Beyond the System
I do not treat AI as a collection of magical black boxes. To build platforms that scale, you must look beyond the models themselves and understand the systems supporting them. For me, AI architecture is software engineering at its highest resolution—requiring the same rigor in state persistence, API contracts, latency budget allocations, and performance optimization as any legacy system.
Staying at the cutting edge requires a continuous feedback loop: reading research papers, analyzing distributed systems designs, exploring new tooling (like the Model Context Protocol or the UV package manager), and writing proof-of-concept (POC) scripts to evaluate performance limitations under load.
I treat AI as a powerful engineering multiplier. I utilize agent runtimes to automate scaffolding, test setups, and documentation. However, I anchor these tools with rigorous context engineering, ensuring structured inputs, predictable constraints, and direct architectural guardrails.
Hands-on POCs first, model hype second.
manoj@research-ledger
STABLE_SYNCResearch & Subscriptions
- Weekly review of arXiv GenAI and System Architecture papers
- Subscribed to engineering publications (Vercel, Stripe, InfoQ)
- LangGraph, LlamaIndex, and OpenAI API spec updates monitoring
System Design & POC Lab
- State preservation using SQLite/Postgres checkpointers
- Benchmarking local SLMs (Qwen-14B/Llama-3.2) on Apple Silicon
- Model Context Protocol (MCP) tool security gateway mocks
AI Integration & Context Eng.
- Utilizing AI for rapid boilerplate generation & testing
- Context engineering: window compaction & RAG compression
- Continuous prompt optimization using structured JSON evaluations
ACTIVE DISCOVERY HUBExplore GitHub
Work With Me
Need an engineering review for your AI system?
Bring the architecture constraint: retrieval quality, agent failure modes, backend latency, evaluation gaps, deployment topology, platform handoff, or developer adoption.