Neeraj Sujan

Work

GTM systems and
AI products I've built.

Engineering background applied to revenue problems. The projects below are real systems — built, shipped, or actively in production.

01

Signal scoring · Data engineering

Open source

GTM Signal Engine

Scores buying patterns, not single events — fewer alerts, every one worth acting on.

The problem

Revenue teams drown in one-off intent alerts: a page visit, a job change, an ad click. Reps stop reading them and miss the accounts that are actually buying.

What I built

  • Medallion pipeline (bronze → silver → gold) on DuckDB and dbt, with weights, decay and combo rules in config files RevOps can tune without code.
  • Pattern multipliers for breadth, channel mix and activity spikes, plus an ICP fit gate, exclusions and a 5-alerts-per-rep daily budget.
  • Dagster runs it daily and pushes "why now" alerts to HubSpot, Attio and Slack, with a Next.js dashboard on top.
PythonDuckDBdbtDagsterNext.jsHubSpotAttioSlack
GitHub ↗
Animated walkthrough: alerts pile up, then data flows through bronze, silver and gold layers into scored alerts.

10/10

planted buyers surfaced as hot

100k

events across 500 accounts scored

50

tests (35 dbt + 15 pytest)

02

AI agent · Sales automation

Open source

Slack SDR Agent

Paste a LinkedIn URL in Slack; 20 seconds later the lead is researched, verified and in the CRM.

The problem

An SDR spends longer researching and logging around each call than on the call itself — and that admin is exactly where follow-ups fall through the cracks.

What I built

  • /prospect command that enriches profile, company, verified email and AI talking points, then upserts HubSpot without overwriting owners.
  • One-click call outcomes: no answer starts LinkedIn + email sequences and finds more buyers; picked up drafts a personalised follow-up for approval.
  • Production ops: idempotent runs, Retry buttons on errors, stuck-work sweeps and a daily summary in Slack.
n8nSupabaseSlackHubSpotProspeoHeyReachInstantlyLLM
GitHub ↗
Slack thread where the SDR Agent posts an enriched lead card with role, verified email, company and research.

~20s

per researched lead (was 10–15 min)

15

n8n workflows, generated from code

33

end-to-end checks, all passing

03

AI SaaS · Founder build

Live · early access

Pipelinewired

An AI agent that finds buyers, writes to each one personally and books the call — built for agencies.

The problem

Agencies stitch together 5–7 tools to run outbound: lead data, signal tracking, sequencer, inbox, dialer, CRM and booking. Nothing ties them into one loop.

What I built

  • Onboarding agent that reads a website and drafts the offer, ICP and personas; a signal-based lead engine scores every prospect before outreach.
  • Multichannel sequences across email, LinkedIn and an AI voice agent, a unified inbox with AI-drafted replies, and booking pages.
  • Agency-native multi-tenancy, a visual workflow builder, compliance controls, Stripe billing and a public API.
Next.jsPostgresAI SDKRetell voiceUnipileStripePlaywright
Live site ↗
Pipelinewired landing page: "Your AI agent finds the buyers and books the call."

40+

product screens shipped

126

test files (unit + Playwright e2e)

6

scheduled background jobs

04

Agentic AI · Prompt chaining

Live

Dhaga Returns Intelligence

Classifies 6,500 weekly return reasons without a data team — three-model chain that routes, synthesises and self-evaluates.

The problem

A fashion e-commerce brand processed 6,500 return reasons a week by hand. The categories were inconsistent, the root causes invisible, and every trend report was already two weeks stale by the time anyone read it.

What I built

  • Haiku classifier that maps each return note to a structured taxonomy in one call; a router that decides whether the case needs synthesis or can be answered directly.
  • Sonnet synthesiser that groups patterns across batches and drafts weekly trend summaries, with a Haiku evaluator that scores confidence and flags low-quality outputs for human review.
  • Streamlit portal with a live returns feed, category breakdown, weekly digest and a prompt-by-prompt trace so ops can see exactly what each model decided and why.
PythonClaude HaikuClaude SonnetStreamlitPrompt chainingAnthropic SDK
GitHub ↗
Dhaga Returns Intelligence: Streamlit dashboard showing return categories, weekly digest and model trace.

6,500

return reasons classified per week

3

model hops per request (classify → synthesise → evaluate)

<2s

per batch classification (Haiku)

05

Azure ML · End-to-end ML system

Live

Contoso Delay Classifier

Predicts e-commerce delivery delays before despatch — trained on 10,000 orders, deployed as a full observability stack.

The problem

A logistics team had no early signal on which orders would miss their delivery window. Delays were discovered reactively, after the SLA breach and after the customer complaint.

What I built

  • GradientBoosting classifier trained on 10,000 synthetic orders with class-balanced weighting; baseline without balancing scored 81.9% accuracy by predicting "on time" for everything — AUC rose from 0.559 to 0.589 once the delayed class was weighted correctly.
  • Azure Functions API that validates inputs, runs inference and writes every request to Cosmos DB with a 30-day TTL; partition key chosen on userId not carrier to prevent hot-partition throughput ceilings.
  • Operations portal with a health panel (p50, p95, error rate over 15 min), recent-requests table and a load test that confirmed cold-start latency of ~1,400ms dropping to sub-400ms on warm instances.
Pythonscikit-learnAzure SQLCosmos DBAzure FunctionsFlaskVercel
Live demo ↗GitHub ↗
Contoso Delay Classifier portal: prediction form, System Health panel with p50/p95 latency and error rate.

0.589

AUC (vs. 0.559 unweighted baseline)

46%

recall on delayed class (was 0% before balancing)

<400ms

p50 latency on warm instance

06

Live product · Content engine

Live

Fieldwork

The hub for Forward Deployed and GTM Engineers: roadmaps, live jobs, posts, people, books and repos.

The problem

FDE and GTM Engineering are new fields. The good material is scattered across LinkedIn, YouTube, podcasts and job boards, buried in hype and recruiter posts.

What I built

  • Ingestion pipeline that builds the catalog from seven sources (Apify scrapers, iTunes, Open Library and curated lists) and never wipes data on a failed run.
  • Quality filters and AI judgments separate real insight from hiring, announcements and promo; people are ranked on insight posts only.
  • Live jobs read from 25 company job boards at request time; a GitHub Action refreshes the catalog daily and Vercel redeploys on every commit.
Next.jsTailwindApifyGitHub ActionsVercelTypeScript
Live site ↗
Fieldwork home page: "Engineers who work in the field," with FDE and GTM Engineering tracks.

25

company job boards read live

160+

curated posts from practitioners

Daily

automated refresh

Announcement card: Neeraj Sujan joins the Founding Cohort of RevTech Academy as a Founding Member.

Credentials & training

Always shipping, always learning from operators.

RevTech Academy ↗ — Founding Cohort

Founding MemberCurrent

Strategy → Context, Process → Architecture, Execution → Orchestration. Partnered with Puzzle and Clay.

Revenue Cartel — Operator Deployment

Sep – Nov 2026

Live 8-week program building six production GTM systems: intent signals, AI research agent, deliverability, personalised outbound, pipeline automation and a capstone.

GTME.SCHOOL ↗ — GTM Engineering program

Sep 2026

ICP definition, TAM/SAM/SOM mapping, buying signals and intent triggers, offer and message-market fit.

Other work

GTMe AI

SaaS ProductIn development

AI-native outbound sales automation. Finds leads, writes personalized outreach, reads replies, and decides the next best action per lead — with a confidence-policy-based approval queue so humans stay in control of the decisions that matter.

Next.js 14SupabaseStripeAnthropic ClaudeInngestTypeScript

GTM Pipeline Tool

Internal ToolBuilt

AI-powered outbound pipeline with two modes: prospect a target company (research → leads → personalized emails), or surface hiring managers at companies actively recruiting GTM Engineers and generate personalized outreach to them. Claude Sonnet handles the research and email generation layer.

Next.jsAnthropic ClaudeApolloHunterTypeScript

GTM Engineering Handbook

Open ResourceGrowing

The field notes, playbooks, and deep dives from neerajsujan.com — organized by topic. Clay workflows, enrichment recipes, outbound system designs, AI agent architectures. Open and searchable.

MDXNext.js

Training

Revenue Cartel — Operator Deployment

Sept 2026 – Nov 2026

Live 8-week operator program building six production-grade GTM machines on real accounts: Intent Signal Engine (Clay, RB2B, Apollo), AI Research Agent (Claude, Claygent), Deliverability System (Smartlead), Hyper-Personalized Outbound sequences, Pipeline Automation Loop (n8n + CRM sync), and a full integrated capstone.

GTME.SCHOOL

Sept 2026

Intensive GTM engineering program covering ICP definition, TAM/SAM/SOM mapping, buying signals and intent triggers, offer and message-market fit, and GTM strategy execution.

Stack

Enrichment & Data

ClayApolloClearbitCoresignalHunterDatagmaDropcontact

Intent & Signals

RB2BKoala6senseBombora

Outbound & Automation

SmartleadEmailBisonHeyReachSalesRobotn8nMake

AI & Engineering

Anthropic ClaudeNext.jsTypeScriptSupabasePostgresStripeInngestPython

Let's talk

Building revenue infrastructure? I'd love to help.

I'm looking for GTM Engineering and Forward Deployed Engineering roles in Switzerland and across the EU. Best fit: B2B SaaS and AI-first teams that want signals, agents and outbound built properly.

Book a 30-min call →Email meLinkedIn