dsrailab is a two-node, local-first research environment. A cloud reasoning brain orchestrates local executor models and MCP tooling over a private tunnel — independent nodes, least configuration, nothing that can't be rebuilt from a repository.
Dinesh Srivastava is an enterprise technology leader and hands-on AI engineer with more than two decades of experience spanning IT strategy, digital transformation, data platforms, cybersecurity, and applied artificial intelligence.
He has led technology at national retail scale across a network of 600+ stores serving Tier 2, 3, and 4 markets in India. At dsrailab, he turns that operating perspective into working systems: production LLM architecture, multi-agent orchestration, MCP integrations, governed data platforms, and secure local-first infrastructure.
His approach is deliberately practical and provider-neutral: define the business outcome, design the controls, build the platform, and verify the behaviour. The work connects long-horizon leadership with the discipline to write, test, and ship the code.
Dinesh has led the implementation, operational management, modernization, upgrading, and automation of enterprise technology across retail stores, the Head Office, distribution centres, warehouses, and hybrid-cloud infrastructure. His remit connects supply-chain execution with data, AI, cybersecurity, identity, and reliable day-to-day operations.
Technology leadership for the major regional distribution hubs at Palwal, Haryana, and Bengaluru, including the approximately 13.60 lakh sq. ft. Palwal facility and roughly 6 lakh sq. ft. of V-Mart-owned or managed warehouse space across the core logistics network.
Implementation, management, upgrades, and automation across Infor WMS and Fretron TMS, strengthening warehouse execution, transport planning, visibility, and operational control.
Leading the ongoing Stackbox WMS orchestration implementation and integration with warehouse control and automation solutions from Falcon Autotech and Armstrong Dematic.
Managing the technology environment serving the Head Office, stores, and more than 700 leaders, managers, and executives, with a focus on service continuity, adoption, governance, and scalable operations.
Oversight of an on-premises data centre and AWS cloud estate, including Data Studio as a hybrid-cloud data lake and Insight Studio for AI-driven collaborative analytics and decision support.
End-to-end security and connectivity across the Head Office, stores, and Hope Box operations, including Fortinet, Active Directory and domain controllers, SSO, multi-factor authentication, and zero-trust networking.
Production LLM and multi-agent architecture, reasoning and planning, tool orchestration, RAG, evaluation, and reliable human-supervised workflows.
MCP server and client design, reusable agent skills, and integration patterns that connect knowledge, code, data, and enterprise workflows.
End-to-end platforms across Postgres, pgvector, and ClickHouse, from ingestion and vectorization through governed analytics.
Postgraduate-certified security practice, threat modelling, privacy-aware architecture, PII isolation, and controls embedded across delivery.
Multi-year technology roadmaps, team and operating-model design, and retail and supply-chain digitization across hybrid estates.
Python, FastAPI, Docker, private local AI runtimes, cloud intelligence services, agent toolchains, vector search, analytical systems, and reproducible infrastructure.
The lab separates reasoning from execution. A cloud model holds the plan and makes the hard calls; small local models do the volume work — code, retrieval, embeddings, drafting — with nothing leaving the network unless a decision genuinely needs the brain.
Everything runs under one rule: the two nodes never depend on each other. Either can be wiped and rebuilt on its own, and the system still stands. What sits on top of that base is a set of specialised programs — coding, knowledge work, design, data, and context — that all share the same router, the same enforced boundary, and the same local-first default.
Running the models next to the data changes what's possible and what's exposed. Iteration stops costing money, latency drops to the network's own speed, and the hardest question in enterprise AI — where did the data go — has a boring answer: it stayed here.
A single hook decides where each task goes. Work passes through a PreToolUse interceptor that classifies it and dispatches to the right backend — reasoning to the cloud brain, everything else to a local executor. A stdio tier-enforcer guards the boundary so nothing silently escalates.
dsrailabintercept.py reads the pending tool call and scores the work against the tier map.FastMCP stdio tier-enforcer holds the boundary across a six-layer harness.Each program is a specialisation with its own crew of models and MCP tools, and each earns its keep the same way — local where it can, cloud where it must, private throughout. Two of them are foundations the rest stand on: a context layer that remembers, and an orchestration layer that decides.
A production-grade, local-first coding system: a cloud intelligence layer handles high-complexity reasoning over a six-layer harness, private local AI handles volume generation and execution, and an elastic cloud specialist supports the hardest work. The cloud layer plans and reviews; the private local layer writes, runs, and iterates; the harness decides what is worth escalating.
An agentic knowledge-work partner that runs briefs, research, drafting, and multi-step tasks against local models first, escalating to the brain only for judgment. MCP connects it to documents, calendars, and tools without handing them to a third party.
An agentic design-and-build loop: turn intent into working interfaces and prototypes, iterate on layout and behaviour with a model in the loop, and hand off clean, reproducible artifacts. Local models generate the volume of variants; the brain arbitrates structure and taste.
A local-first data agent that models schemas, wires pipelines, and answers questions over a warehouse — natural language to SQL, semantic layers, and quality checks — with embeddings and retrieval held inside the perimeter.
The orchestration layer that composes the crews — routing a goal across coding, knowledge, design, and data agents, holding shared state, and choosing which specialist and which tier each step needs. Not a claim of general intelligence: a general-purpose coordinator over specialised, verifiable parts.
A knowledge hub that turns a personal vault and its graph into live context — connected through MCP to the agents, versioned in Git, synced with the tools already in use — with search and embeddings served by local models.
The shared context substrate beneath everything — a private memory and knowledge-graph layer of retrieval, embeddings, and relationships that gives every agent long-lived, relevant context instead of a cold start each session. Local by default, so context accrues without ever leaving the network.
A browser-based, mobile-responsive order-management platform for Indian retail. It brings marketplace, ONDC, webstore, store/POS, warehouse, logistics and ERP workflows into one operational model, with deterministic routing and an advisory AI copilot.
dsr-OMS normalises orders from marketplaces, ONDC, the brand webstore, stores and warehouses into one canonical workflow. Operators see orders, shipments, routing, inventory, locations, integrations, AI insights and audit history in a single responsive workspace.
The transactional core is deliberately deterministic. It reserves sellable stock, minimises shipment splits, selects eligible fulfilment locations, enforces a four-hour allocation-acceptance SLA, and releases stock before rerouting when a location does not accept on time.
Central views for multi-channel orders, shipments, exceptions, inventory, locations, integrations and audit history.
Scores sellable stock, safety stock, service zone, proximity, location priority and warehouse bias while minimising splits.
Confirm, GST invoice, delivery partner and AWB, pick, pack, manifest and dispatch with guarded state transitions.
Return request, item-level QC, sellable or damaged disposition, restock handling and refund queueing.
Signed idempotent webhooks plus a transactional Ginesys invoice/refund outbox with retry, replay and reconciliation.
Channel catalogues, master data, PIN codes, logistics, ERP, gateway policy, RBAC and platform controls without storing secrets in the UI.
Local models turn iteration into a free operation. Large agent loops, evaluation runs, and adversarial tests all execute on-machine, so behaviour is validated at volume without a meter running — the difference between reading about an agent's failure mode and reproducing it a thousand times before it reaches anything real.
MCP is the standard socket. Every tool, dataset, and vault plugs into the agents through one protocol, so new capability composes instead of being re-integrated each time. The same harness that runs experiments runs production, which means what you validate is what ships — new behaviour is prototyped on local models, gated through the enforcer, and promoted to the brain only where judgment is genuinely required.
The default is closed. Content leaves the perimeter only when a decision genuinely needs the brain, and that path is explicit — an agent cannot quietly route sensitive work to the cloud, because the tier-enforcer holds the boundary rather than trusting the model to respect it.
Credentials are the usual leak. The brain authenticates over OAuth with no API keys written to disk, removing a whole class of exposure; the nodes reach each other over a private WireGuard mesh behind a Cloudflare Tunnel and are never published directly to the internet.
Independence is the resilience story. The two nodes share nothing, so either can fail without taking the system with it, and recovery is a repository checkout rather than a backup restore. Because every service is reproducible, the worst case is measured in minutes.
Reached over a WireGuard mesh and a Cloudflare Tunnel, published at dsrailab.com.