Open to Forward Deployed Engineer opportunities
Nikhil Rajput · Forward Deployed Engineer
I build AI products that move from conversation to production.
I take messy business workflows from first discovery into production AI agents, automation systems, full-stack platforms, and live integrations — engineered to survive real users and real operational constraints.
- Doing this
- Nearly 3 years, hands on
- Based in
- Madhya Pradesh, India
- Scope
- Discovery → production
Fig. 1One request, all the way through
100+development applications captured dailydata acquisition · Australia
20+council portals tracked overnightdata acquisition · Australia
250+property listings ingested monthlyPropTech SaaS · South Africa
<500msto first audio on live PSTN callsvoice AI · India
16+safety guardrails on every model calllegal intake · Australia
2-Tierlocal element + vision action verificationShiv AI · Android & TV
0-LLMcalls on verified repeat workflow replaysWebPilot AI & Shiv AI
1,000+students managed per academic sessioneducational ERP · India
What I bring
The intelligence layer is the easy half.Everything it touches is the job.
My best work happens in the overlap — where AI, full-stack engineering, integrations and somebody's actual working day all have to agree with each other.
I’m a Forward Deployed Engineer with nearly 3 years of hands-on experience taking AI products, autonomous agents, automation systems, full-stack platforms, and production integrations from messy client discovery to live deployment.
I work from the first confused conversation through architecture, build, testing, deployment and the long tail of changes that follow. I'd rather carry that line end to end than hand off half of it.
Read my complete résuméI start where the job actually happens
Not the one in the process document. I map the people, the spreadsheets, the WhatsApp group where the real decisions happen, and the edge cases nobody mentions until week three.
I own the connected build
Interfaces, APIs, databases, agents, workers, integrations, tests and the deploy pipeline. One person holding all of it means fewer seams for things to fall through.
I design for the day it misbehaves
Structured outputs, validation, bounded tools, retries, fallbacks, logging and a clean human handoff. Not features bolted on later — the shape of the thing from the start.
What I've shipped
Shipped systems, with theengineering story attached.
Client delivery across AI, automation, data acquisition and SaaS — then my own products, where I get to make the architecture calls and live with them.
Client systems · 06 engagements
Delivered for teams in South Africa, Australia, India, and under NDA — each one replacing work that used to happen by hand, in a spreadsheet, or not at all.
Multi-Tenant Construction & Field Service Operations Platform
One multi-tenant backend behind web and mobile apps, built to keep working offline.
A multi-tenant operations platform for construction and field-service teams, in active development, with web and offline-capable mobile clients on one shared backend. I own backend architecture and database design, and build the platform's AI and third-party integrations.
- Backend engineering
Modular monolith with GraphQL and REST APIs on NestJS and TypeScript, shared typed contracts, strict validation, and RFC 9457 error handling across clients.
- Database design
Multi-tenant PostgreSQL schema in Prisma with tenant isolation, versioned migrations, and concurrency-safe writes.
- Access control
Role-based permissions checked on every server action rather than trusted to the client.
- Offline sync & events
Idempotent, ordered event sync for offline field devices with reliable publishing over Terraform-provisioned AWS SNS/SQS queues.
- Human-approved AI
Wired AWS Bedrock LLMs, Textract OCR, and vector search off the request path so AI outputs arrive as suggestions a person approves.
Multi-Tenant AI PropTech SaaS
Listings, tenant ops, owner outreach, and social distribution unified on one system.
Replaced disconnected listing spreadsheets and manual property outreach with an isolated multi-tenant platform where acquisition, outreach, syndication, and action-taking WhatsApp agents share one live data layer.
- Data harvesting
Session-managed Playwright workers sanitising and ingesting 250+ rental listings monthly from diverse portals into an isolated multi-tenant database.
- Owner outreach
Automated communications engine that messages property owners as soon as tenant interest arrives and extracts intent from replies.
- Social syndication
n8n workflows generating AI-written listing copy and publishing to Facebook Marketplace, Pages, and Instagram, cutting manual entry by over 90%.
- Action-oriented agents
WhatsApp RAG bot and in-app agent wired to bounded backend commands so users can list properties or raise service requests in natural language.
AI Legal Intake & Safety Infrastructure
An SMS intake agent that qualifies claimants without ever giving legal advice.
Scoped strict legal compliance boundaries into an SMS qualification platform on FastAPI and Twilio — collecting facts, scoring readiness, and escalating sensitive cases to lawyers under Australian data residency.
- Controlled intake
Configurable SMS conversation flows with structured fact extraction, lead scoring, and lawyer escalation built on FastAPI and Twilio.
- Safety & guardrails
Inbound and outbound safety gates with crisis detection, restricted-topic rules, and output validation across 16+ documented guardrails.
- Deterministic core
State-machine orchestration combining rule-based qualification, explicit scoring logic, and fallback model routing.
- AU residency & audit
Pinned Bedrock inference to Australian regions with an append-only conversation audit log and separate guardrail-event trail feeding HubSpot.
Conversational Voice AI & Marketing Agents
Live PSTN calls under two seconds, handed to WhatsApp or a person with full context.
A real-time telephony voice stack for delivery partner onboarding and education enquiries that handles live calls, carries state into WhatsApp, and absorbs the front-line work of four to five marketing executives.
- Low-latency streaming
Bidirectional FastAPI audio pipeline built to a strict turn budget (~150ms STT, 400–600ms model, 200–400ms to first audio) with barge-in and per-turn instrumentation.
- Multi-provider fallback
Resilient adapters across Gemini, Groq, Claude, Sarvam, and Whisper so live calls survive upstream provider slowdowns.
- Cross-channel continuity
Central state store moving callers seamlessly from a phone call into a WhatsApp RAG assistant without losing context.
- Live warm handoff
Function-calling intent triggers executing instant transfers to human agents alongside HubSpot sync and a Next.js ops dashboard.
Enterprise Data Acquisition Platform
Twenty council portals watched overnight, turned into enriched CRM leads by morning.
Eliminated daily manual portal checks by monitoring public planning sites overnight, extracting fields from planning PDFs, enriching corporate records, and routing qualified leads to sales.
- Distributed scraping
Resilient Playwright workers tracking 20+ Australian council portals and capturing 100+ new development applications daily.
- Document extraction
Async AWS Bedrock and Textract pipelines pulling builders, applicant contacts, and project scope from unstructured planning documents.
- Automated enrichment
Real-time ABR ABN Lookup and RP Data checks adding corporate contacts and property context to every lead.
- CRM sync
React monitoring console wired to Monday.com with instant client email triggers, cutting manual routing and data entry by 95%.
Construction Slab Layout & Take-Off Platform
Draw the slab, place the pods, sync measurements straight into the site's engineering sheet.
A construction planning tool that computes exact void-former pod layouts from raft-slab plans and reads/writes quantities directly against the engineering team's existing Google Sheet.
- Slab geometry
Parses multi-slab plans with Shapely and normalises boundaries onto the true pod setout grid.
- Exact pod placement
Places full field pods and perimeter fractional pods checked against real polygon containment rather than bounding boxes.
- Google Sheets sync
Writes slab thickness, pod height, area, and rebates into the engineering sheet and reads calculated quantities back.
- Delivery
Shipped as a Flask, MongoDB, JWT, and React workspace generating Plotly setout drawings for sign-off.
Integration & automation
The tools a business already runs, finally talking to each other.
CRM, inbox, WhatsApp, voice, LinkedIn, Drive and Sheets are the ones already wired — the orchestration layer takes anything with an API, deterministic where it matters, with agents only on the language-shaped steps.
Cross-Tool Workflow Automation Suite
One automation layer wired across CRM, inbox, WhatsApp, voice, LinkedIn, and Sheets.An n8n and Node.js integration layer joining the tools a business already runs — keeping deterministic logic in code and restricting AI strictly to unstructured language steps.
CRM sync
Two-way HubSpot sync across contacts, deals, and activity from web forms, LinkedIn, and inbound calls without manual retyping.
Messaging & voice
WhatsApp Business, Brevo email, and voice calling wired together so replies, call outcomes, and opt-outs land back on the CRM.
Browser-native outreach
LinkedIn Chrome extension running inside the user's own browser with profile-aware messages, pacing delays, follow-ups, and A/B variants.
Invoice pipelines
Webhook, Gmail, and Drive triggers fenced by schema validation, duplicate checks, and retry queues before writing to Sheets or CRM.
Safety & control
Bounded LLM steps with deterministic routing, failure alerts, and human approval wherever actions are hard to undo.
My own products
Where I make the architecture calls and live with them.
Independent product work — evaluation, safety boundaries, async systems and the interface all move together.
Shiv AI — Autonomous Android & TV Agent
Voice-first autonomous OS control across any Android phone or TV app — verified step by step.A native Kotlin agent that listens offline for a wake phrase, drives third-party apps via Android Accessibility, verifies every screen step before moving on, and turns repeated tasks into fast local replays.
- Offline wake & live voice: Local Vosk wake detection keeps mic audio on-device until triggered, then streams full-duplex voice to Gemini Live with instant barge-in.
- Verified execution & replay: Confirms each UI step via fast local element checks before vision fallback, and freezes completed runs into zero-LLM macros that re-learn if an app shifts.
- Safety outside the model: Gates payments, calls, and wipes against the user's own spoken words, masks OTPs and card numbers, and auto-pauses during UPI or banking flows.
WebPilot AI
Teach a browser workflow once, replay it with zero model calls, and heal it when the site changes.Live across 15 Terraform-managed Google Cloud services, turning plain-language goals into approval-gated browser workflows that run deterministically until the page changes.
- Zero-model fast path: Gemini learns a live browser workflow once behind human approval, then every healthy repeat run replays the frozen spec via Playwright with zero reasoning calls.
- Verified self-healing: Breakages trigger DOM/screenshot diagnosis, sandbox patching, independent second-model verification, and immutable version promotion before resuming.
AI Career & Job Intelligence Platform
An AI-first operating system for the entire job search, not another resume tool.Career profiles, job ingestion, vector matching, tailored applications, and spoken interview prep unified in one asynchronous product stack.
- Scheduled connectors for LinkedIn, Indeed, Naukri, and Wellfound with deduplication, skill extraction, and Weaviate vector matching.
- Redis and Taskiq workers running parsing, matching, resume generation, and alerts off the live request path.
- LiteLLM multi-provider routing and LangGraph workflows traced end to end in LangSmith, plus voice interview practice.
Financial Analytics AI Platform
Static bank statements turned into a queryable, auditable financial ledger.Parses 100+ page PDF bank statements into clean relational rows, then answers questions grounded strictly in verified ledger data.
- FastAPI extraction pipeline turning 100+ page PDF bank statements into validated PostgreSQL transactions.
- Hybrid retrieval combining pgvector semantic search with generated SQL so answers resolve against real rows, not summaries.
- React analytics dashboard across 1,000+ historical transactions, cutting manual auditing turnaround by around 85%.
Additional products
More proof, across more categories.
A compact view of the wider full-stack, integration, mobile, language and geospatial systems I've worked on.
Omnichannel MarTech SaaS
Publishing and paid campaigns across five networks from one console.A centralised product for multi-platform OAuth connections, scheduled publishing, and unified paid ad operations across major networks.
- OAuth 2.0 authorisations and token-refresh lifecycles across 5+ platforms — Meta, TikTok, YouTube, LinkedIn, and X.
- Unified campaign creation and spend reconciliation spanning Meta, TikTok, Google, LinkedIn, and X Ads.
Enterprise Educational ERP
Full student lifecycle for 1,000+ students an intake, replacing legacy check-in hardware.An institutional ERP unifying admissions, GPS-geofenced QR attendance, PhonePe fee flows, WhatsApp updates, and placements.
- Launched end-to-end student operations across admissions, attendance, payments, and placements for 1,000+ students per session.
- Replaced legacy hardware with GPS-geofenced QR check-ins while supporting 20+ faculty on daily academic ops.
Real Estate CRM Bridge
An external CRM normalised into a multi-tenant schema, live on web and mobile.Keeps listings, contacts, and enquiries synchronised between a third-party real-estate CRM, operational dashboards, and an Expo app.
- Built a FastAPI bridge normalising deeply nested external CRM JSON into a clean multi-tenant PostgreSQL schema.
- Structured agency onboarding via sync scripts with live Firebase updates pushed to React and Expo mobile clients.
Field Interview Transcription Pipeline
Hour-long Hindi field interviews turned into schema-validated agricultural records.An audio-to-structured-data pipeline converting noisy rural recordings into clean records ready for downstream analytics.
- Combined Google Cloud Hindi Speech-to-Text with Gemini extraction across hour-long field recordings.
- Enforced crop, soil, and practice schemas with Pydantic, routing low-confidence fields to human review.
Precision Agriculture Platform
Nine satellite analysis layers for any farm boundary drawn on a map.A geospatial platform turning remote-sensing imagery into zone-wise farm diagnostics and temporal NPK trends.
- Generated nine analysis maps per field from Sentinel Hub alongside optional ISRO VEDAS India overlays.
- Built an OpenLayers + FastAPI workspace where drawing a farm boundary computes NPK trends and prescription maps.
Every system above shipped. Ask me what broke first — that story is usually more useful.
How I work
Useful AI startslong before the model call.
The model is one component. The workflow, the data it's allowed to see, the controls around it and the recovery path decide whether anyone keeps using the thing in month two.
four habits, in order
- 01no architecture on day one
Sit with the process before the stack
Before choosing a model or a framework, I map who does the job today, what they actually touch, where the handoffs break, and what the business is really paying for.
- 02the leash is the design
Draw the line between rules and reasoning
Money, eligibility, compliance and anything irreversible stay deterministic. The model gets the messy, language-shaped middle — and a short leash of allowed tools.
- 03one problem, not seven
Build the full path, not the clever part
Interfaces, APIs, workers, queues, databases, integrations, logging and deploys are one problem. A great agent behind a broken webhook is still a broken product.
- 04assumptions lose
Ship it into the working week
I test the ugly paths, release in slices, watch what production does to my assumptions, then change the system with the people using it.
Technical range
Grouped by the part of the systemthey hold up.
I pick the stack around the workflow, the deployment environment and the reliability bar — not around whatever I used last time.
FDE & AI Agents
Backend & Mobile
Frontend
Automation & Voice
Data
Cloud & DevOps
Integrations
How the responsibility grew
Quality first. Then software.Now complete AI systems.
It explains how I work now: I still hunt edge cases like a tester, build like an engineer, and keep the product in view the entire time.
Three roles in nearly three years, each one deeper into ownership.
Forward Deployed Engineer
I lead forward-deployed AI and full-stack delivery — from client discovery and architecture through agent workflows, integrations, guardrails, deployment, and production iteration.
- AI agents, RAG pipelines, structured extraction, bounded tool calling, safety guardrails, verification, and human-in-the-loop workflows
- Production systems with Python, FastAPI, NestJS, Node.js, Kotlin, PostgreSQL, n8n, Playwright, Redis, AWS, GCP, and third-party APIs
Software Engineer
I built backend and full-stack applications for international clients across Australia, South Africa, and other markets — turning manual workflows and fragmented data into production software.
- Node.js, Express.js, Python, FastAPI, React, Next.js, PostgreSQL, and MySQL
- REST APIs, webhooks, CRM integrations, dashboards, scraping pipelines, deployment, and production support
QA Tester + Developer
I started close to product quality: testing real workflows, hunting edge cases, and contributing fixes and debugging alongside developers.
- Functional, API, regression, and end-to-end workflow testing
- Defect reproduction, edge-case validation, fix verification, and cross-team debugging
Where I fit next
Give me a real problem,not an isolated demo.
I'm looking for a team where engineering ownership, AI capability and actual business context sit in the same room.
Agents that bridge frontier models and real client production
Forward Deployed Engineer · AI Solutions Engineer · Applied AI Engineer
The complete product, integrations, and guardrails wrapped around the model
Agentic AI Engineer · Full-Stack AI Engineer · AI Product Engineer
Business friction on Monday, verifiable systems shipped into production
AI Automation Engineer · Integration Engineer · Product-minded Engineer
What a team gets
I don't stop at“the AI part.”
I can sit with the people doing the job, turn a vague complaint into a decision, and keep the engineering pointed at the outcome instead of the ticket.
System boundaries and schema design through to APIs, agents, dashboards, automation, integrations and the deploy. I go all the way down.
I know when plain code beats a model, where AI genuinely earns its cost, and where validation, fallback, observability or a human has to be in the loop.
Start a conversation
Got a process thateats everyone's week?
I'm in Madhya Pradesh, India, open to Forward Deployed Engineer, Agentic AI, Full-Stack AI, automation and AI solutions roles, and happy to work with teams in other timezones. Tell me what your team is trying to make faster, cheaper or less painful.
