Backend systems & APIs
REST services, auth and roles, async workers and database-first design built to hold up under real traffic, not demo traffic.
DjangoFastAPIDRFCeleryWebSocketsPostgreSQL
Available for freelance engagements
I don't start with AI. I start with the workflow.
I'm Yashvardhan — a freelance backend systems & AI engineer based in India. I build deterministic production architectures (Django, FastAPI, PostgreSQL, Redis) paired with grounded AI automation (PGVector RAG, LLM guardrails, automated decision mapping). I eliminate repetitive manual operations from team workflows — then prove business impact with exact, tracked operational metrics.
Deterministic Core
Grounded RAG
Zero-Trust FinOps
High-Throughput APIs
0%P99 latency
reduction
reduction
0%Infrastructure
cost cut
cost cut
0%Support tickets
deflected
deflected
0Production
deployments
deployments
Workflow triage
01 / 04
Is the work repetitive and rule-based?
Deterministic
Automation, not AI
EngagementAutomation Sprint
How I decide
Not every workflow needs a model.
Most of what arrives labelled as an AI problem is a workflow problem wearing a costume. I diagnose the process first, then pick the cheapest thing that holds: rules where rules work, models where the input is genuinely messy, and a person in the loop wherever a wrong answer is expensive to undo.
AI automation
Retrieval-grounded assistants, document extraction and lead qualification — with a prompt firewall in front and a strict schema behind.
RAGPGVectorGeminiOpenAIClaudeGuardrails
Custom software
Internal tools, dashboards, CRMs and admin systems — the unglamorous software a business actually runs its day on.
ReactNext.jsTailwindAnalyticsKanban CRM
Cloud, DevOps & performance
Infrastructure as code, CI/CD with tests and scans on every push, caching strategy and latency work — so releases become boring.
AWSDockerTerraformNginxGitHub ActionsZero-trust VPC
AI Pipeline & Guardrail Architecture
01. INGEST
User Input
WebSocket stream over Django Channels
→
02. FIREWALL
Prompt Guard
Regex & heuristic sanitizer blocking injection
→
03. RETRIEVAL
PGVector RAG
Embedding lookup with 0.82 similarity floor
→
04. INFERENCE
Gemini / Claude
Grounded LLM output with JSON schema
→
05. PERSIST
CRM Sync
Entity extraction & PostgreSQL auto-lead scoring
PILLAR 01
Revenue & High-Concurrency Scale
Eliminating database locks, N+1 query bottlenecks, and server crashes during flash sales and peak marketing surges.
26.6x Peak Capacity BoostPILLAR 02
FinOps & Infrastructure Optimization
Cutting cloud waste, right-sizing databases, and implementing edge caching to lower monthly AWS bill by 60%+.
68.2% Infra Cost CutPILLAR 03
Automated Workflows & AI Deflection
Deploying retrieval-grounded AI assistants and prompt firewalls to resolve repetitive customer inquiries with 0 hallucinations.
84.2% Support DeflectedPILLAR 04
Zero-Downtime & Uptime Protection
Containerized Docker pipelines, Terraform IaC, automated rollbacks, and zero-trust security for 99.99% system availability.
99.99% Uptime SLAThe stack, by layer
What I reach for, and why.
Frontend
ReactNext.jsTailwind CSSTypeScriptISR
Fast, responsive UI with performance-first delivery
Backend
Django 5FastAPIDRFFlaskCeleryChannels
Clean APIs, async workloads, systems that scale sideways
AI
GeminiOpenAIClaudeRAGPGVectorPrompt firewall
Grounded output, checkable against its own sources
Data
PostgreSQL 15Read replicasMongoDBRedis 7SQL tuning
Reliable storage, caching and queues under load
Infrastructure
AWSDockerTerraformNginxGunicornCI/CD
Reproducible deployments and disciplined rollouts
In development
What I'm building right now.
In development · Confidential engagement · Phase 1 · 2026
Operational intelligence & decision mapping platform
Lead architect & engineer · LLM pipeline, evidence lineage, analyst review
An operations-consulting client needed to turn qualitative organisational evidence — survey answers, interview notes, policy descriptions — into an auditable map of who actually decides what. That synthesis normally costs weeks of manual consulting. I'm building the pipeline that does it: extract atomic claims from free text, resolve everyone's different vocabulary onto one canonical taxonomy, infer decision ownership and required inputs, then hand an analyst a review workspace before anything ships. The model proposes, the evidence supports, a human still governs.
- Signed-webhook intake bridging an existing CMS front end into a FastAPI backend
- Immutable evidence layer — every output traces back to the original wording, source and timestamp
- Three-tier normalisation: alias dictionary → pgvector similarity search → LLM reasoner, with low-confidence matches auto-flagged
- Contradiction and ambiguity detection that refuses to pick a winner and escalates to a human instead
- Versioned assessments, so new evidence supersedes rather than overwrites
- Prompt-injection containment: untrusted text isolated, extraction model given zero tool access, scoped-JWT RBAC
Web intake signed webhook
│
├─ Evidence immutable responses + provenance
│
├─ Extraction LLM → atomic claims, strict JSON
│
├─ Normalise alias → vectors → reasoner
│
├─ Synthesis decisions · owners · inputs · limits
│ └─ conflicts flagged, never guessed
│
├─ Review analyst approves / annotates
│
└─ Export multi-tab decision matrix
└─ every cell traceable to a quote
7Pipeline stages
3-tierSemantic normalisation
100%Evidence lineage
FastAPIPostgreSQL + pgvector
- Architecture, data model & pipeline specComplete
- Intake gateway & evidence storeIn progress
- Extraction & normalisation enginesNext
- Review workspace & matrix exportPlanned
Selected work
Systems already in production.
Live · Enterprise platform · 2026
BlueShore Technologies platform
Built as a freelance engineer for Blueshore Technologies (blueshoretech.com)
Blueshore is a B2B software and digital growth agency. They needed their own operating system: a CRM, a visitor-intelligence layer and an AI assistant that could capture and qualify leads without inventing answers. Sales signals were scattered across separate tools; I built the single platform that replaced them, end to end.
- Kanban CRM with proposals, invoices and contract modules
- Gemini assistant streaming over WebSockets, grounded by a RAG pipeline
- Real-time visitor telemetry feeding live analytics dashboards
- SEO/AEO middleware with Redis caching; SSL, Celery workers, hardened defaults
Users / Sales team
│
├─ CRM Leads → Deals → Proposals → Invoices
│
├─ Assistant Gemini + RAG (human escalation)
│ └─ capture · qualify · summarise
│
└─ Analytics Visitor events → dashboards
└─ real-time telemetry
99.8%+Uptime SLA
3×Faster delivery
Django 5DRF + Channels
Redis+ Celery workers
Live · Brand & retail · Fashion
London Berry Clothing
Brand website & collections experience
A contemporary menswear and womenswear house founded in 2016, with 200+ exclusive outlets across India. The brief was reach, not e-commerce: tell the brand story, open clean entry points into each collection, and let a shopper find the nearest store in two taps — on a phone, on a weak connection.
- Conversion-oriented structure: brand story → collections → store discovery
- Component-based UI with reusable sections for future campaigns
- Performance-first delivery tuned for Core Web Vitals on mobile
- SEO-ready content layout and a safe, monitoring-friendly release setup
Brand story
│
├─ Collections menswear · womenswear
│ └─ reusable campaign sections
│
├─ Discovery search · filter · lookbook
│
└─ Store finder 200+ outlets, pan-India
└─ two taps to nearest store
200+Exclusive outlets
2016Brand founded
Core WebVitals tuned
Pan-IndiaRetail footprint
Own product · FinOps · Open source
CloudPulse
Cloud governance & cost intelligence
A tool that finds the AWS resources nobody is using and says what to do about them. FastAPI and Next.js, AI-driven cost recommendations, and a zero-trust IAM model so the scanner can read your account without ever being able to change it.
cloudpulse-chi.vercel.app →~30%Cloud spend cut
Zero-trustIAM model
FastAPI+ Next.js
DockerDeployment
Inside the platform
The interface, not a mockup.
Screens from the Blueshore Technologies platform I built end to end — the operations dashboard, the CRM, the marketing site and the AI assistant. Click any screen to open it full size.
Architecture, documented
Every system ships with its map.
Handover documentation from the same platform. If the next engineer can't read how it works, it isn't finished.
A single request traced end to end — Nginx terminating SSL and rate limiting, Django routing through middleware to business logic, PostgreSQL and Redis behind it, and the async paths where Celery workers and WebSocket broadcasts branch off.
Click to enlarge
The assistant pipeline: prompt construction and rate limiting in Django, streaming responses from Gemini over server-sent events, entity extraction writing scored leads into the CRM — and the local fallback engine that keeps answering when the model API times out.
Click to enlarge
The PostgreSQL schema: UUID primary keys throughout, soft-delete state flags, unique SEO slugs, and cascade rules chosen table by table — CMS, CRM, careers and audit logging in one relational model.
Click to enlargeEngineering case studies
Four engagements, measured the same way.
Pick the metric the business already feels. Move it. Keep the receipt. Delivered as a freelance engineer through Blueshore Technologies for their enterprise clients.
P99 API Latency
93.3% Speedup
Before
4,200ms
After
280ms
Monthly Infra Cost
68.2% Cost Cut
Before
$14,500
After
$4,600
Checkout Capacity
26.6x Growth
Legacy
450/m
Scaled
12,000/m
Deployment SLA
97.7% Faster
Manual
180 min
CI/CD
4 min
Support Response SLA
99.9% Faster
Human
14 Hours
AI Agent
< 1.2s
Mobile PageSpeed
+188% Boost
Monolith
34 / 100
Headless
98 / 100
Case study 01
Global fintech SaaS platform redesign
A legacy monolith on Python 2 buckling under 300% year-over-year signup growth. N+1 queries pushed response times past four seconds at market open, synchronous transaction processing blocked Gunicorn workers into 504s, and there was no centralised audit logging to survive a PCI-DSS review. I rebuilt the backend on Django 5 with read-replica routing, a Redis caching layer, and Celery workers handling invoice generation and webhook dispatch off the request path.
ClientCross-border payments provider
Scale1.2M monthly active users
Timeline14 weeks, two phases
RoleFull-stack engineer
P99 API latency
4,200 ms→280 ms
93.3% faster, against a sub-300ms SLA target
Monthly infrastructure cost
$14,500→$4,600
68.2% saved — roughly $119k a year
System uptime
98.1%→99.99%
Zero unplanned downtime after cutover
Case study 02
Autonomous AI support & lead qualification
5,000 inbound tickets a month, a 14-hour first response, and 42% of high-value leads leaving before anyone called them back. This is the case where a model earns its place — but only behind guardrails. Input arrives over Django Channels with per-IP throttling, passes a prompt firewall that rejects injection attempts, then hits a PGVector retrieval step with a 0.82 similarity floor before Gemini is allowed to answer. Anything below the floor escalates to a person.
ClientB2B logistics & supply chain group
Volume5,000 inquiries / month
Timeline6-week sprint
RoleFull-stack / AI integration engineer
First response time
14 hours→under 1.2s
Instant on every channel, including weekends
Ticket deflection
0%→84.2%
~4,200 tickets a month resolved without a human
Qualified lead volume
110/mo→460/mo
4.18× growth, auto-scored HOT / WARM / COLD into the CRM
Annual support cost
$220k→$40k
$180,000 saved per year
Case study 03
Omnichannel e-commerce & high-traffic portal
Every seasonal campaign ended the same way: inventory row locks during concurrent checkout, cascading crashes, and thousands of orders accepted for stock that no longer existed. I split the storefront into a Next.js head over Django REST catalog services, pre-rendered 95% of catalog pages with ISR so browsing never touches the database, and moved stock reservation into atomic Redis Lua scripts. Reads spread across three PostgreSQL replicas; the primary now does nothing but write.
ClientGlobal retail & consumer goods
Peak traffic25,000 concurrent shoppers
Timeline10 weeks
RoleFull-stack engineer
Lighthouse, mobile
34→98
+188% on the score shoppers actually feel
Largest contentful paint
5.8s→1.2s
79.3% faster via CDN, WebP/AVIF and edge caching
Peak concurrent checkouts
450/min→12,000/min
26.6× capacity — and zero over-selling
Campaign revenue
+340% YoY
Record sales peak, no incident during the window
Case study 04
Multi-region DevOps & zero-trust security
Multi-tenant healthcare data spread across manually configured servers, three hours of scheduled downtime per release, and config drift quietly making staging lie about production. I codified every environment in Terraform, moved production databases and Redis into private VPC subnets with no route to the internet, and put a GitHub Actions pipeline in front of every push: 50+ tests, ClamAV scanning, container build, automated rollback on failure.
ClientHealthcare data management platform
ComplianceHIPAA · ISO 27001 · SOC 2 Type II
DeploymentMulti-region Docker clusters
RoleDevOps & backend engineer
Deployment duration
180 min→4 min
Manual maintenance window replaced by a pipeline
Deployment frequency
Every 2 weeks→Daily+
Small releases, small blast radius
Change failure rate
14.2%→under 0.1%
Tests and scans gate every merge
Mean time to recovery
4.5 hours→under 2 min
Automated rollback, no human in the recovery path
Business Impact & ROI Simulator
Estimate cost savings & support deflection for your scale
$122.8K
Annual Cloud Savings
4,210 / mo
Support Tickets Deflected
$176.8K
Annual Support Savings
< 1.2s
AI Response SLA
Who you'd be working with
One engineer, end to end.
I'm Yashvardhan Mishra. I started in enterprise delivery at Accenture, spent a year leading web application builds at Adviwe Media, and went independent in early 2026. Since then I've shipped CRMs, RAG assistants and cloud infrastructure for agencies and their enterprise clients.
Working with a freelancer usually means a handoff problem — one person scopes, another builds, nobody owns the deployment. I don't split it up. The person who writes your architecture doc is the person who takes the 2am page, which is a useful incentive for getting it right the first time.
Since 2024Building in production
Django · FastAPICore specialisation
IST · flexibleOverlaps EU & US mornings
BCA, 2024Computer studies, Kanpur
"Send me the workflow, not the AI idea." — How every project starts
What clients say
The part I can't measure myself.
"Yash was instrumental in bringing Blueshore's platform vision to life — from the CRM and AI assistant to the analytics layer. He communicated clearly, delivered on schedule, and consistently anticipated technical decisions before we had to raise them.
"He understood the brand before he wrote a line of code. The site is quick on a phone, the collections finally look the way they do in store, and customers can find their nearest outlet without hunting for it.
How an engagement runs
Four steps, in order.
01
Days 1–3
Technical scoping
I review your current architecture, the pain points, and the business metrics we're aiming at.
02
Week 1
Blueprint & audit
A technical roadmap with fixed milestones and performance SLAs written down before anyone codes.
03
Week 2+
Sprint execution
Two-week sprints with continuous Docker staging access and automated test reporting.
04
Handover
Production & support
Zero-downtime deployment, documentation you can hand to the next engineer, and post-launch support.
Track record
Where I've built.
Feb 2026 — Present
Freelance full-stack & AI engineer
Self-employed · Remote
Production SaaS in Python, Django, FastAPI, PostgreSQL, Redis and Docker. Enterprise chatbots, RAG systems, CRM platforms and workflow automation — engaged directly by agencies and their enterprise clients, from requirements through deployment and tuning. Currently lead architect and engineer on an LLM-based operational intelligence platform.
Mar 2026 — Jul 2026
Junior Software Engineer
C-Zentrix · Gurugram
Backend systems and AI chatbot solutions. Built scalable Python APIs for production and shipped features quickly inside a cross-functional startup team.
Feb 2025 — Mar 2026
Software Engineer
Adviwe Media · Remote
Led end-to-end development of scalable web applications. Designed REST APIs, database schemas and automation workflows; mentored developers and ran code reviews.
Sep 2024 — Feb 2025
Packaged App Development Associate
Accenture · Gurugram
Python REST APIs and automated deployment pipelines that cut release time by 30%. Integrated AWS EC2, S3 and IAM; resolved 50+ critical production issues.
2021 — 2024
BCA, Computer Studies
Dr. Virendra Swarup Institute · Kanpur
Certifications: AWS APAC Solutions Architecture (Forage) · Foundations of UX Design · Software Design & Development Methodologies · Responsive Web Design.
Credentials & Learning
Academic foundation & verified certifications
AWS Solutions Architect
APAC Architecture · Forage
Foundations of UX Design
Google Design Certificate
Software Design & Dev
Architecture & Methodologies
BCA Computer Studies
Dr. Virendra Swarup Inst (2021-24)
Next step
Have a workflow that eats hours every week?
Send me the workflow — not the AI idea. I'll tell you what should be automated, where a model actually earns its place, and what it would take to build. No charge for that conversation.
© 2026 Yashvardhan Mishra · India — working remote
Backend systems · AI automation · Cloud