Production-Deployed · Lead AI Developer Portfolio

I Built an Autonomous
AI Hiring Engine

Upload a CV. Press Search. Watch jobs appear from 15+ portals — ranked, explained, and CV-optimised — with applications auto-submitted and contacts found. End-to-end, zero manual steps.

∞
Jobs ingested
scales without limit
18
Pipeline stages
fully automated
15+
AI agents
orchestrated
Live
SSE streaming
to the browser
🐍 Python · FastAPI 🤖 GPT-4o · pgvector ⚡ BullMQ · Redis SSE 🔭 OpenTelemetry 📦 Docker · Railway 🎯 Adaptive ML weights
✉ Request Repository Access
The Problem

Job search is a data pipeline
nobody should run by hand

Hundreds of portals. Duplicate listings. Generic CVs. No ranking. No feedback loop. This is an engineering problem masquerading as a life problem — so I automated it.

01
Fragmented signal
Relevant roles scattered across 15+ portals with no unified ranking or deduplication. Hours of manual triage every day.
02
Generic CVs get rejected
ATS systems discard non-tailored CVs before a human reads them. No scoring, no iteration — just silence.
03
Zero observability
No data on which skills are in demand, which portals convert, or which applications lead to interviews. Flying blind.
The Product — Live Search

15+ portals scraped in parallel.
Every job ranked against your CV.

One click triggers concurrent scraping across LinkedIn, Indeed, Jobber, Bundesagentur, Xing, Stepstone, Greenhouse, Lever, Ashby and more. Real-time progress streams to the browser via Redis SSE — no polling, no refresh.

Live · Redis SSE

What happens behind
the scenes

Playwright launches headless browsers and navigates each portal. For every listing it extracts the job URL, title, company, and full description — then inserts directly into Supabase, deduped by SHA-256 hash.

  • Playwright fetches job URL + full description from each portal page
  • Raw HTML normalised to Markdown before storage
  • SHA-256 content hash prevents duplicate inserts across portals
  • Jobs upserted into Supabase · embeddings computed via OpenAI
  • INGEST → RANK phases streamed live — progress visible in real time
15+ Portals · Parallel Scrape
Live search progress dashboard
The Product — Intelligent Ranking

Every job. Every dimension filterable.

Multi-factor scoring surfaces the right roles. Filter by match score tier, application status, role, location, and source portal — all computed server-side via hybrid pgvector + FTS retrieval.

Hybrid pgvector + FTS

Role-aware ranking
across every portal

Not keyword search — semantic vector similarity fused with full-text scoring and adaptive per-user weights. 194 jobs ranked by how well they match your actual CV.

  • Score tiers: ≥85% · 70–84% · <70% · Unscored
  • Status: Active · PDF Ready · Applied
  • Role breakdown: AI Engineer 68 · Data Eng 62 · Software Eng 57
  • Sources: LinkedIn 63 · Jobber 62 · LinkedIn-MCP 38 · Indeed 25
  • Sort: Best Match · Similarity % · Date Posted
Multi-portal · Hybrid Ranking
Smart filter panel with role and source breakdown
The Product — AI Document Generation

One click. Tailored CV + Cover Letter.
Compiled to PDF.

Hit Generate on any job card. GPT-4o scores, rewrites, and iterates your LaTeX CV until it clears the ATS threshold — then compiles both CV and cover letter to downloadable PDFs. Multiple jobs run in parallel.

4
Pipeline steps
Score → Optimize → Cover → PDF
94
ATS match score
achieved post-optimisation
∞
Parallel jobs
BullMQ corner bar
3
Optimisation loops
average to hit threshold
Parallel generation · Live progress
Generate corner bar with parallel CV generation
Corner bar — multiple jobs generating simultaneously, real-time progress per task
Real CV output · 2 pages · PDF
Generated CV PDF output
ATS-optimised CV for Autom SRL — compiled by tectonic LaTeX engine, stored in Supabase
The Product — Market Intelligence

Real-time labour market data.
No third-party subscription.

Every scraping run feeds an analytics layer — skill demand charts, role distribution, tooling trends, and a GPT-4o generated personalised learning roadmap. Data that normally costs thousands per year.

Aggregated · Live

Skill demand, role mix,
and tooling trends

The market page aggregates across all ingested jobs. No manual data entry — it's a by-product of the pipeline running normally.

  • Average AI match score computed across all scored roles
  • Python #1 demanded skill — 69% of all postings
  • AI Engineer top role — 8 open positions tracked
  • Top tools: AWS · Airflow · Snowflake · TensorFlow · Kubernetes
  • GPT-4o personalised learning roadmap — one click generate
AI-scored · GPT-4o Insights
Market intelligence dashboard with skill demand charts
The Product — Network Discovery

Hiring contacts found.
Automatically.

The contact finder agent runs for every saved job — discovering real LinkedIn profiles at the hiring company. Talent acquisition specialists, HR leads, recruiters — surfaced with one click.

LinkedIn · Enrichment agent

Real people at
real target companies

Contacts at Schwarz Digits include SAP SuccessFactors consultants and talent specialists — exactly the right people for a SAP-adjacent role. Found and stored automatically, no manual searching.

  • Contacts stored per company with LinkedIn URLs — grows with every save
  • Schwarz Digits: SAP SuccessFactors consultants and TA specialists found
  • TAWO: Talent Acquisition Manager discovered and stored
  • Contact finder agent runs async via BullMQ
  • Direct "View Profile" links for immediate outreach
LinkedIn enriched · Auto-discovered
Connections page with real LinkedIn contacts
The Product — Application CRM

Every application tracked automatically.
Every status. Every company.

Every auto-submitted application lands in the CRM — company, portal, score, date applied, and CRM status. QuantumBlack (McKinsey), Axel Springer, Bending Spoons — all tracked without a single manual entry.

Auto-tracked · Gmail sync

Application lifecycle
from submit to offer

The pipeline logs every application the moment it's submitted. Gmail OAuth syncs replies back as status transitions — Awaiting Response → Interview → Offer — with zero manual entry.

  • Every submitted application auto-logged — grows with every run
  • Companies: QuantumBlack/McKinsey · Axel Springer · Bending Spoons · acemate.ai
  • Status stages: Applied · Interview · Offer · Rejected · Ghosted · Withdrawn
  • Sort by: Date Applied · Score · Company
  • Filter by: All · Applied · Interview · Offer · Rejected · Ghosted
Auto-tracked · Gmail sync
Applied Jobs CRM with 121 applications
Architecture

Three services. One coherent system.

Every boundary is intentional — independently deployable, separately scalable, connected through typed contracts and a shared secret.

Browser Gateway + Dashboard
Next.js 16 App Router — Port 3000
All user traffic enters here. Route Handlers enforce auth via httpOnly JWT cookies + Supabase RLS. Real-time search progress streamed via SSE. React 19 · TanStack Query · TailwindCSS 4 · Zustand.
↓ HTTP (X-Internal-Secret)    BullMQ queues
Python ML Compute — Port 8000
FastAPI + uvicorn
Owns all ML ops: embedding (OpenAI text-embedding-3-small), hybrid retrieval (pgvector ANN + FTS), 18-stage pipeline, deterministic scoring, adaptive weights, Cohere reranking, GPT-4o explainability. 59KB pipeline.py.
Background Workers — Node.js
BullMQ Consumers (cv-tasks · search-pipeline)
worker.ts handles score_job, generate_cv, generate_cover_letter, apply_job. search-pipeline-worker.ts delegates full scrape+rank runs to FastAPI. Both publish Redis pub/sub → SSE → browser in real time.
↓ psycopg3 pool    ioredis
Persistence
Supabase Postgres + pgvector
10 RLS-protected tables. HNSW vector index + GIN full-text index. 12 ordered migration files. Service-role key for workers (bypasses RLS); anon key for browser (enforces RLS). Zero privilege escalation possible.
Queue + Realtime Cache
Redis — BullMQ · SSE · Cancellation
BullMQ queues with retry + exponential backoff. Per-search progress pub/sub. LLM explanation cache (24h TTL). Cancel flag propagation — aborts in-flight pipeline runs without orphaned processes.
↓ OTLP exporter
Observability Stack
OpenTelemetry → Jaeger · Sentry · Langfuse
FastAPI + all httpx outbound calls auto-instrumented. OTLP span export to Jaeger. Sentry on Next.js for frontend error capture. Langfuse hook for LLM prompt versioning and call tracing. Per-stage timing in search_trace.py.
End-to-End Pipeline — SA-00 → SA-18

18 stages. Zero manual steps.

From user clicking Search All to ranked, explained, CV-optimised matches with optional auto-apply — fully automated, observable, and cancellable at any point.

01
Parallel Scraping
15+ portals fetched concurrently. Bundesagentur · LinkedIn · Jobber · Indeed · Xing · Stepstone · Greenhouse · Lever · Ashby · Crunchbase.
02–03
Dedup + Embed
SHA-256 hash dedup across portals. HTML → Markdown. Batch embed via OpenAI text-embedding-3-small into pgvector HNSW.
04–05
Hybrid Retrieval + Score
pgvector cosine ANN fused with Postgres FTS. Multi-factor composite score: semantic · role tags · location · recency · title.
06–07
Rerank + Explain
Optional Cohere/cross-encoder reranking with wall-clock budget guard. GPT-4o match explanations cached 24h in Redis.
08–10
Optimise → Apply
LaTeX CV rewritten by GPT-4o until ATS score ≥ threshold. Compiled to PDF via tectonic. Playwright auto-fills and submits.
SAP Lead AI Developer — Role Alignment

Every requirement demonstrated
in production code

Not claimed on a CV — implemented, deployed, and observable. Each capability maps directly to the SAP job description.

🐍
Advanced Python · FastAPI
Required · Backend frameworks
59KB async pipeline.py orchestrating 18 stages. Fully typed with Pydantic, psycopg3 connection pool, 50+ validated env vars, startup health checks, CORS and internal-secret middleware.
FastAPIasyncioPydanticpsycopg3
🤖
Multi-Agent Architecture
Nice-to-have · LangChain, CrewAI
15 numbered agent modules (00 vector-store → 24 startup-funding-discovery), each with typed I/O contracts. Orchestrated via BullMQ — decoupled, retryable, observable. GPT-4o feedback loops with safety guardrails.
GPT-4oBullMQ15 Agents
🔄
MLOps / LLMOps
Required · CI/CD, continuous training
Adaptive weight system derives per-user ranking deltas from behavioural signals — zero-downtime retraining. CV scorer gates every application. Pipeline guardrails with wall-clock budget enforcement per stage.
Adaptive MLGuardrailsLLMOps
📦
Docker + CI/CD
Required · Docker, Kubernetes, GitHub Actions
Multi-stage Dockerfile bundles Node 20 + Python venv + tectonic LaTeX + Playwright Chromium into one reproducible image. docker-compose for full local stack. Railway CI/CD for production.
DockerComposeRailwayMulti-stage
🔭
Observability
Required · Reliability, compliance of AI solutions
OpenTelemetry SDK auto-instruments FastAPI and all httpx calls — traces export to Jaeger. Sentry on Next.js. Langfuse for LLM prompt versioning. Structured pino logging with per-stage timing.
OpenTelemetryJaegerLangfuseSentry
🗄️
Data Engineering + Vector DB
Required · ETL, data modelling, distributed platforms
ETL: HTML → Markdown → embed → dedup → upsert. pgvector HNSW + FTS hybrid retrieval. 12 ordered migration files — schema as code. RLS on every table. Service-role isolation for workers.
pgvectorHNSWRLSETL
Full Stack

Technology Inventory

Every tool chosen for a reason — not for résumé padding.

AI / ML
🧠 GPT-4o — scoring · CV optimisation · cover letters · explainability 📐 text-embedding-3-small — HNSW ANN retrieval 🔁 Cohere · Cross-Encoder — optional second-stage reranking 📊 Adaptive weights — SA-18 behavioural feedback loop
Backend
🐍 FastAPI + uvicorn 🔗 psycopg3 connection pool 📋 Pydantic v2 schemas 🌐 Next.js 16 App Router Route Handlers ⚡ BullMQ + Redis queues 📡 SSE pub/sub streaming
Data + Storage
🐘 Supabase Postgres 🔍 pgvector HNSW + GIN FTS hybrid 🔐 Row-Level Security 📂 12 migration files — schema as code ⚡ Redis — cache · queues · SSE
Frontend
⚛️ React 19 + Next.js 16 App Router 🎨 TailwindCSS 4 🔄 TanStack Query 5 📦 Zustand 5
Infrastructure + Observability
🐳 Docker multi-stage + Compose 🚂 Railway CI/CD 🔭 OpenTelemetry → Jaeger 🐛 Sentry 📊 Langfuse LLM tracing 🎭 Playwright browser pool 📄 tectonic + pdflatex
Private Repository

Want to see the code?

The repository is private — but I'll add you as a read collaborator within the hour. Full access to commit history, open PRs, every file and migration. No tours, no filtered demos. Just the real codebase.

✉ nihildev.nandakumar@gmail.com

Or reply directly — response within the day.