UNKAPPED
Founding forward-deployed engineer at an applied AI and GTM engineering firm: inside client teams in cybersecurity, DevTools, cloud and AI-native software, taking what they bought into production.
Forward-deployed engineer
You bought the AI. The number still hasn’t moved.
That’s the gap I work in. I sit inside a client’s team, in the systems they already pay for, and turn the tools they bought into systems that run in production. Then I stay until they run without me.
- Founding FDE at UNKAPPED
- Runs Xeme
- AI GTM engineer since 2020
- Official Clay Creator
- 3x Clay Certified
- 25+ systems in production
- Pune, works remote
Running now
Meeting intelligence, live at UNKAPPED.
Every recorded outside call becomes one record tied to the right client, with a summary a person reads first. Nothing reaches a client or the CRM without a gate.
Drawn from the live workflow map, structure only5 capture doors · one record per call
n8n / the real workflow29 workflows · 29 live · 1,620 tests · 27 Oct 2026
Every recorded call, one record a person trusts.
- 1Every capture door saves each recording once.
- 2Recordings of the same call merge into one: 179 transcripts became 109 calls.
- 3Unsure calls go to a person, marked unsure.
- 4Ask answers only from the call text and quotes it: 22 asks, 16 answered.
- 5Tasks go to one board only, capped at 15 new a day. Nothing is written to HubSpot.
- 6Loud error alerts and a token keeper watch all 29, every one switched on.
29n8n workflows
29live now
1,620tests pass
179 → 109transcripts to calls
- Zero-retention AI
- 12-month purge
- No-AI clients never processed
- Client names only in the database
- No transcript kept in run history
- Code de-identified, private repo
Case study: one record per call
The problem
Calls came in from 7 recorders, and the same call often arrived more than once. 179 transcripts turned out to be 109 calls.
What runs
Every recording is saved once, merged into one call, tagged by type, matched to the client in HubSpot (read only) and summarised. Calls the system is unsure about go to a person.
In production
All 29 workflows live, 9 of them switched on today. In the last 7 days 638 of 653 runs finished clean. 47 summaries across 39 calls, and 22 questions put to the call library, 16 answered with quotes. 1,620 tests pass.
Switched on today
Task drafts into one team board (at most 15 new and 30 changed tasks a day, confidence floor 0.6), review buttons, a pre-call brief, the pilot digest, calendar gaps, Gemini notes and the daily health check. Nothing is written to HubSpot.
Read from the live n8n and the database on 27 Oct 2026.
Where I work
One day job, one practice, one notebook.
And one product I am building alongside the client work.
Xeme
Applied AI GTM for AI startups from Seed to Series B, with fixed prices: Teardown $1.5K, Sprint $15K, Lab from $7.5K a month.
After the Demo
My notebook on Substack: what breaks when AI goes to production, one outcome at a time.
Dokeo
The product I am building alongside the client work.
Systems in production
Anyone can build a workflow. These keep running.
A workflow needs someone to watch it. A system has a schedule, retries, an alert when it goes quiet, and tests that fail loudly.
Pre-discovery brief engine
Inbound qualification from 4 hours to under 3 minutes, with a branded brief on the AE’s desk before the first call. 8 n8n workflows, a 5-provider enrichment waterfall and a B2B gate; 698 unit and 143 end-to-end tests. Moved into UNKAPPED’s own n8n on 27 Oct: 9 workflows published, first test runs in.
n8n, Claude Code, Supabase
Meeting intelligence, live
179 transcripts became 109 calls, each tagged, matched to the client and summarised for a person to review in one click. Account memory, call scores and search sit on top. 29 workflows, all live, 1,620 tests passing.
n8n, Supabase, Claude, Slack, HubSpot
Relay, reply triage
Claude Haiku classifies every LinkedIn and email reply and drafts an answer. A person decides, and nothing sends on its own. Multi-tenant Supabase with row-level security, 183 commits in 16 days.
Supabase, Claude, HeyReach, Smartlead, MCP
Seven-stage copy engine
Role map, research, signal, gap, angle and a draft under 90 words, then a second model checks facts and spam risk. 1.5%+ positive replies across 150+ domains.
Clay, Claude Code, Smartlead
ABM pages
500+ pages, one per account, built on whether AI assistants named that company in its own category. Every visit tracked to the page view.
Ahrefs, OpenAI, Claude, Supabase
Visitor to sequence
Anonymous traffic resolved to a person and an account, enriched, scored and in a sequence the same day.
RB2B, n8n, Clay, HubSpot
AI lead scoring
Lead qualification from 48 hours to 9.6, with written reasons pushed into the CRM. Part of $180K ARR growth in six months.
Clay, Python, HubSpot
Qualification gate
Scores people across several intent sources, rejects anything it cannot confirm and records why: 102 of 1,562 got through. Rejecting is the feature.
Python, Supabase, Modal
Xeme's own engine
942 contacts enriched, 584 verified emails and 242 of 275 buyer emails passing the copy check, synced back to HubSpot with a reason.
Clay, n8n, HubSpot, Postgres
Automated Inbound
Demo requests routed to the right rep with a first-call brief. Free checks catch test, duplicate and invalid entries before any credit is spent, and paid enrichment runs only behind gates. Passed 23 Oct 2026, graded by Clay.
Clay, HubSpot
Outbound Automation
A messy 71-row CRM export became three validated, ready leads. Excluded accounts are never contacted, and every paid step sits behind a gate. Passed 24 Oct 2026, graded by Clay.
Clay, CRM export, sending gate
Paid Ads, the hard one
A 601-company LinkedIn audience, no personal data. Clay matched 175 to a LinkedIn page; I opened every one and 29 were wrong, so they were held out. First attempt scored 11 of 25, then 18 submissions: 23 of 25 and 25 of 25 on the interview, 28 Oct 2026.
Clay, LinkedIn Ads
How I work
GTM treated like a software system.
Aim first: TAM, ICP, personas, the signals worth chasing. Then build what executes it.
Speed
If an experiment takes a week, rebuild the process until it takes days.
Always on
Repeated GTM work becomes a system that keeps running.
Zero friction
Seven tools and two approvals to ship one campaign? Remove the handoffs.
Learning loops
Find the bottleneck, test one idea, ship, measure, repeat.
Rules I keep arriving at
Fail closed
When the data is uncertain, skip the person and record why. Most tooling fails open and quietly contacts the wrong people.
Signal before volume
Sending more is not the answer. Two intent sources agreeing beats one source shouting.
Silence lies
A quiet morning looks like success. Alert on zero, not only on errors.
Own it, do not rent it
If the handover is a recording and a spreadsheet, the buyer rented a workflow.
The record
GTM systems, built in-house since 2020.
Every number below comes from the role it sits under.
Sep 2026 to now
Founding Forward Deployed Engineer
UNKAPPED, applied AI and GTM engineering firm, San Francisco. Remote.
- Works inside client teams in cybersecurity, DevTools, cloud and AI-native software, taking the AI and GTM tools they already bought into production.
- Pre-discovery brief engine: 8 n8n workflows, a 5-provider enrichment waterfall, a B2B gate and a branded PDF brief. 698 unit and 143 end-to-end tests. Inbound qualification from 4 hours to under 3 minutes. Moved into UNKAPPED's own n8n on 27 Oct: 9 workflows published, first test runs in.
- Clay outbound engine with a credit cap and a QA gate: 53 tests, 1,200+ ready-to-send copies across 3 segments, sent from Clay to PlusVibe by API across 300+ mailboxes.
- Meeting intelligence, live: recordings from Google Meet, Granola, Fireflies, tl;dv and a notes form are saved once and merged (179 transcripts became 109 calls), tagged by call type, matched to the client in HubSpot and summarised for one-click review; unsure calls go to a person. Account memory, call scores and search over calls are built on top; task drafts and HubSpot write-back stay off until approved. 29 workflows, all live, 1,620 tests passing, zero-retention AI.
- The team's GTM playbook as a web edition, A4 PDF, Claude skill and Slack bot.
Aug to Sep 2026
Senior GTM Engineer
OneGTMLab, on the DefenderMate engagement (CTEM security). Remote.
- ICP and scoring model, Clearcue signal watchers, a Reclaim to HubSpot meeting gate and an RB2B visitor webhook.
- Relay, a reply triage engine: LinkedIn and email replies stored once in Supabase with row-level security and an audit log; Claude Haiku classifies intent and drafts a reply; a person decides. Nothing sends on its own.
- 16 Edge Functions, 34+ migrations, an MCP connector for the client's own Claude, 183 commits in 16 days.
Oct 2025 to Aug 2026
AI GTM Engineer
Omnibound AI, AI search marketing platform, Dallas. Remote.
- Revenue engineering as a one-person function. Campaign deploy time from 2 weeks to under 3 days.
- A 7-stage LLM personalization pipeline, agentic account research, an RB2B visitor engine and 500+ programmatic ABM pages built on AI share-of-voice data, every visit tracked.
- 95%+ deliverability and 1.5%+ positive replies across 150+ domains and 300+ mailboxes.
Apr 2023 to Jun 2025
BDR to AE to Founder's Office GTM Lead
RemoteState, software development firm. Noida.
- Full-cycle AE: $550K closed-won, $2M BANT-qualified pipeline, $3.5M total pipeline. Promoted from Senior BDR after four months.
- GTM Lead in the Founder's Office: 3,000+ accounts scored with Clay and a Python scorer, a Clay to HubSpot pipeline, lead qualification from 48 hours to 9.6.
- $180K ARR growth in six months, with weekly tests on copy, timing and signal type.
Sep 2024 to now
Clay Creator
Clay, official creator program. Remote.
- Enrichment waterfalls, ICP scoring, signal routing and CRM sync, with credit caps and QA gates.
- 3x Clay Certified: Automated Inbound (23 Oct 2026) and Outbound Automation (24 Oct 2026), each with a 23 of 25 build and a 25 of 25 interview, and Paid ads (28 Oct 2026).
Dec 2022 to Apr 2023
Fractional Head of Growth
FalconWise Tech, cybersecurity and third-party risk. Dubai, hybrid.
- ABM engine from zero pipeline across 5 data sources: 10+ SQLs a month at 30% engagement, targeting accuracy up 45%.
- A co-marketing channel with 3 regional vendors that brought 25% of pipeline.
Jun to Dec 2022
Senior BDR, APAC
CareerLabs, EdTech. Pune.
- Outbound rebuilt on Reply and HubSpot: SQL conversion from 16% to 27%, 50+ demos at a 65% show rate.
2020 to 2022
Earlier
The foundation years.
- Business development at Skill-Lync (lead scoring on HubSpot, research time per lead down 50%). Relevel business development program. BCA, Computer Science, Sabarmati University.
Writing
Field notes from real builds.
n8n said “Succeeded”. Nothing had been saved.
A green tick, a workflow that finished, and the part I cared about never ran. Why I now check which steps actually ran, not the tick.
13,700+ followers
Shorter notes from the same work: forward-deployed engineering, applied AI and GTM engineering.
Kit
What I build with.
If your stack already has some of these, the build starts from what you have.
01
AI and automation
- Claude Code
- Codex
- Anthropic API
- OpenAI API
- Gemini
- MCP servers
- OpenRouter
- n8n
- Modal
- Apify
- Firecrawl
02
Signals
- Clearcue
- Trigify
- PhantomBuster
- TheirStack
- Common Room
- Bombora
- RB2B
03
Outbound and enrichment
- Clay
- Sales Navigator
- Smartlead
- Instantly
- HeyReach
- Findymail
- ZeroBounce
04
CRM and calls
- HubSpot
- Salesforce
- Attio
- Pipedrive
- Fathom
- Sendspark
05
Data and AI search
- Python
- Playwright
- SQL
- Supabase
- PostgreSQL
- Vercel
- GA4
- Ahrefs
- Semrush
- Mixpanel
06
Certified, October 2026
- Clay Certified: Automated Inbound
- Clay Certified: Outbound Automation
- Clay Certified: Paid ads
- HubSpot Revenue Operations
- HubSpot Sales Hub Software
- HubSpot Marketing Hub Software
- HubSpot Inbound
Next step
A pilot that worked in the demo and stalled after?
Tell me where it stalled. That’s the conversation I want to have.
Or ask an assistant about me:
ChatGPT
Claude
Google AI Mode
Perplexity
Grok