The 13-Month Pitch — Six Versions of One Sentence¶

Who you are on Monday, July 5, 2027. Rehearsed starting M11 (May 2027). Delivered from M12 onward.

Everything in this fortress compiles into one sentence. The rest is decoration for that sentence. Below, the sentence is expanded to six versions for the six contexts you will need it in — from a 30-second elevator to a portfolio landing page to a LinkedIn bio to the awkward “why should we hire you” beat in a technical study.

Each version is grounded in specific shipped artifacts. If the sentence is not backed by the rung, you cannot use the sentence. That’s the discipline.

The Core Sentence¶

“I write C that ships in production — from SIMD-accelerated ML inference kernels to multi-threaded network services — and I can debug someone else’s segfault at 2 AM without flinching.”

Memorize this. Say it in the shower. Say it before you fall asleep. By M13 it should come out of your mouth in one breath without hesitation.

Why each clause is defensible¶

Clause	The rung that makes it true
“I write C that ships in production”	Rung 7 — merged non-trivial PR into Valkey / llama.cpp / curl / nginx
“SIMD-accelerated ML inference kernels”	Rung 6 — SIMD (AVX2/NEON) GEMM within 3x of OpenBLAS, roofline plot, 2000-word blog
“multi-threaded network services”	Rung 5 — epoll HTTP/1.1 server at 100k+ RPS, 24h leak-free soak
“debug someone else’s segfault at 2 AM”	Rung 7 + Rung 4 — real merged PR + `nsh` shell (Unix process model, signals, fork/exec)
“without flinching”	Rung 1 — the toolchain-journal repo: 5 documented gdb sessions, 5 documented ASan catches

If you have all seven rungs, you can say the sentence. If a rung is missing, drop the corresponding clause — do not fake it.

Version 1 — The 30-Second Elevator (12-15 seconds actual)¶

Context: someone asks “so what do you do?” at a Bangalore meetup, at a wedding, at a chai stand outside the office. You have ~15 seconds before their eyes drift.

“I’m an Applied ML engineer at Zoho, and for the last year I’ve been going deep into C — wrote a 100k-RPS network server and a SIMD matmul that gets within 3x of OpenBLAS. So I do ML on top and C on the bottom. Trying to move fully into ML infrastructure.”

Notes:

Names the role, names the depth, names the intent.
Two concrete numbers (100k RPS, 3x OpenBLAS) make it un-fake-able.
Ends with a direction, not a destination. “Trying to move” is honest and inviting.
Cut it here. Do not add “and I also read Modern C by Gustedt.” That kills the pitch.

Version 2 — The 2-Minute Recruiter Call¶

Context: a recruiter at Cloudflare / Postman / Neon / Sarvam AI messages you on LinkedIn. “Tell me about yourself.” You have two minutes.

“I’ve been an Applied ML engineer at Zoho for the last three years, working on <specific ML product/pipeline>. About a year ago I realized that most of my recent work had been LLM-assisted — the code was shipping but I couldn’t debug it under load. So I put myself on a 13-month plan to relearn systems from the ground up.

“Since then I’ve shipped seven public artifacts: a container library in C, all 75 NeetCode problems in C, a POSIX mini-shell, an epoll HTTP server sustaining 100k RPS with a 24-hour clean soak, and a SIMD GEMM that gets within 3x of OpenBLAS on my machine. The latest one is a merged PR into Valkey — .

“I’m looking to combine the ML side and the systems side — ideally on an ML infrastructure or inference team. That’s what I’ve been building toward, and it’s what I want to spend the next years doing.”

Notes:

Opens with the atrophy story. Recruiters love a self-awareness beat.
Middle is specific and un-fake-able.
Ends with a directional ask. Recruiters need to know where to slot you.
90-110 seconds spoken. Practice it.

Version 3 — The 5-Minute Technical Intro¶

Context: the technical part of an on-site or a real technical study. The study partner says “walk me through your background” and expects depth.

“I’ll do this in three parts — where I came from, what I built in the last year, and what I’m curious about next.

“Background: I’ve been at Zoho for three years, applied ML — working on <specific pipeline / feature>. In the last year of that I started noticing that my day-to-day work had drifted almost entirely to LLM-assisted coding. The code was shipping and my metrics looked fine, but I was losing the ability to reason about performance and about correctness under adversarial input. So a year ago I made a deliberate plan to relearn systems from the ground up.

“The last year: I gave myself thirteen months and seven public deliverables. First was a toolchain-journal repo — basically documented gdb and AddressSanitizer sessions, because I wanted to re-earn the right to my terminal. Then a container library in C — vector, hashmap, heap, deque — with 90%+ test coverage. Then all 75 NeetCode problems in idiomatic C. Then a POSIX mini-shell with fork/exec, pipes, redirects, signal handling. Then — this is the one I’m proudest of — a single-threaded epoll HTTP/1.1 server. It sustains 100k+ RPS on my 4-core laptop under wrk, and I ran a 24-hour soak that came out clean. TSan clean too. Then a SIMD GEMM — I picked NEON since I’m on an M-series Mac — that gets within 3x of OpenBLAS on 1024x1024 fp32. I have a roofline plot and a 2000-word writeup for that. Most recently, a merged PR into Valkey — <specific issue / feature>.

“What I’m curious about next: I want to combine the ML side and the systems side. What excites me is the ML inference stack — quantization kernels, low-precision GEMM, throughput on real production traffic. I’ve read llama.cpp end-to-end and contributed a small change; I’ve read valkey’s networking layer. I want to work somewhere that lives in that intersection. That’s why I’m talking to you.”

Notes:

Structure: past / present / future. Three beats.
Rehearse it in front of a mirror. Time it. Cut anything that puts it over 5 minutes.
The atrophy admission is a strength, not a weakness. It reads as engineer-with-taste rather than engineer-with-résumé-inflation.
Never end with “and I’m learning X.” Always end with “and I want to work on Y with people like you.”

Version 4 — The Portfolio Landing Page¶

Context: github.com/<yourhandle> README. The first thing a recruiter sees when they click through your GitHub. First 5 seconds are decisive.

# Hi, I'm <Name>.

Applied ML engineer at Zoho, based in <Bangalore/Chennai>, moving toward
ML infrastructure and systems.

I write C that ships in production — from SIMD-accelerated ML inference kernels
to multi-threaded network services — and I can debug someone else's segfault
at 2 AM without flinching.

## The last 13 months, in seven artifacts

1. **c-toolchain-journal** — gdb + AddressSanitizer field notes
2. **libprep** — C container library (vector, hashmap, heap, deque) — 90%+ coverage, ASan clean
3. **neetcode-75-in-c** — all 75 problems, idiomatic C
4. **nsh** — POSIX mini-shell (~2000 LOC): fork/exec, pipes, redirects, signals, jobs, history, tab completion
5. **epoll-http-server** — 100k+ RPS on a 4-core laptop, 24h soak clean, TSan clean
6. **simd-gemm-benchmark** — NEON GEMM within 3x of OpenBLAS, roofline analysis, 2000-word writeup
7. **valkey#<PR>** — merged production contribution

## What I'm looking for

ML infrastructure or systems roles — Bangalore, Chennai, or remote.
Places where C and ML both matter: inference engines, kernels,
vector databases, network services close to the ML stack.

## Reach me

- <email>
- LinkedIn: <url>
- Blog: <url>

Notes:

Above the fold: name, current role, one-sentence pitch, seven links. Everything else is scroll-down.
Each artifact link goes to the actual repo. No 404s. This is table stakes.
“Reach me” is not optional. Recruiters do not chase.

Version 5 — The LinkedIn Headline + Bio¶

Context: LinkedIn. Headline is 220 characters. Bio (“About” section) is 2000. Both must be scannable.

Headline¶

ML infrastructure & systems C · Applied ML @ Zoho · Shipped: 100k-RPS epoll server, SIMD GEMM (3x of OpenBLAS), Valkey contributor · Bangalore/Chennai

About¶

Applied ML engineer at Zoho for three years. Over the last year I put myself on a deliberate plan to relearn C from the ground up — not out of nostalgia, but because I wanted to work on the ML infrastructure that sits under the models, not just the models.

The result is seven shipped artifacts on GitHub:

• A POSIX mini-shell (~<phone_number_or_numberic_id_or_random_id_10> LOC) with fork/exec, pipes, redirects, signals, jobs, history, and tab completion • A single-threaded epoll HTTP/1.1 server sustaining 100k+ RPS on a 4-core laptop, 24-hour soak clean, TSan clean under worker-thread mode • A SIMD-accelerated matrix-multiply benchmark landing within 3x of OpenBLAS on my machine, with roofline analysis • A merged non-trivial PR into Valkey / llama.cpp / curl / nginx (as applicable) • And the four smaller rungs that make those defensible: a container library, all 75 NeetCode problems in idiomatic C, an ownership-annotated linked list library, and a documented toolchain journal.

What I’m looking for: ML infrastructure or systems roles where C and ML both matter. Inference engines, quantization kernels, vector databases, network services close to the ML stack. Bangalore, Chennai, or remote.

Not looking for: pure application-layer ML work, or C work disconnected from ML. The whole point of the last year was to combine them.

Blog: · GitHub:

Notes:

“Not looking for” is important. Filters out the noise. Recruiters respect specificity.
Concrete numbers in the About section. Not adjectives.
Update the headline every 3-6 weeks with your latest ship. LinkedIn algorithms notice.

Version 6 — The “Why Should We Hire You” Answer¶

Context: the moment in an study when they ask, directly or indirectly, “why you?” You have 90 seconds. You will be tempted to be humble; do not be. This is not the time.

“Three reasons.

“One: you’re looking for someone who can move between ML and systems. I’ve spent three years doing applied ML at Zoho, and the last thirteen months rebuilding my C fluency to the point where I can ship a 100k-RPS server and a SIMD matmul that competes with OpenBLAS. Very few candidates you’ll see have both sides live at once. I do.

“Two: I know what I don’t know, and I’ve proven I can close the gap on a calendar. A year ago I couldn’t have debugged a race condition under wrk. Now I have a public repo where I did exactly that. If your team needs to onboard someone onto , I have a track record of doing that on my own time, in the open, on a schedule.

“Three: I want this specific kind of work. I’ve read llama.cpp’s ggml, I’ve read valkey’s networking layer, I’ve written a roofline analysis of my own kernel. This is what I want to spend my next years on. You will get an engineer whose curiosity aligns with the actual work you have.

“That’s the pitch.”

Notes:

Three reasons. Structured, memorable, un-mumbleable.
Reason 2 is the self-awareness beat again. It repeats because it works.
End with “that’s the pitch.” It is a closer, not an opening for a follow-up ramble.
Practice this one hard. It is the single most important 90 seconds of the M13 hunt.

What You Are Explicitly NOT Claiming¶

The discipline of the pitch is what you do not say. Never claim these:

“C expert” or “systems expert.” Say “systems C” if needed. Never “expert.” You are one shipped rung better than most junior candidates, one shipped rung short of most seniors. Own the middle.
“Fluent in kernel development.” You touched drivers in Phase 6 as intro. That’s not fluency.
“Deep knowledge of GPU / CUDA.” Explicitly out of scope. Do not fake it. Say “I focused on CPU inference this cycle.”
“Contributions to the Linux kernel.” Unless Rung 7 was a kernel patch (unlikely by M13), do not claim this.
“Contributor to ” when the contribution was a typo fix. Rung 7 requires ≥200 LOC or documented significance. Anything less is not “contributor”; it’s “one-line-fixer.”
“Rewrote X in C.” Nothing you did in 13 months qualifies as a rewrite. You wrote from scratch. Say that.
“Best C programmer at Zoho.” Do not benchmark yourself against colleagues in a pitch. Pitch is about you, not about them.

If you catch yourself drifting into any of the above during a call, stop mid-sentence and correct. “Sorry, I misspoke — what I meant was…” That correction is worth more than the pitch itself.

Rehearsal Schedule (Starting M11)¶

Rehearsal begins the moment Rung 6 ships — not before. You cannot rehearse a pitch built on unfinished evidence.

Week (2027)	Rehearsal activity
May 10-16 (S23-w1)	Write all six versions in a doc. Do not memorize yet. Sleep on it.
May 17-23 (S23-w2)	Read all six out loud to yourself. Cut ruthlessly. Nothing over the word counts.
May 24-30 (S24-w1)	Record V1 (elevator) and V2 (recruiter) into your phone. Listen back. Cringe. Rewrite.
May 31 - Jun 6	Q4 BUFFER week — rest. Do not rehearse.
Jun 7-13 (S25-w1)	Deliver V1 to one non-technical friend. Deliver V2 to one recruiter (Pramp counts).
Jun 14-20 (S25-w2)	Rehearse V3 (5-min technical) 3x during the week. Time it. Cut anything over 5 min.
Jun 21-27 (S26-w1)	Publish V4 (portfolio landing) and V5 (LinkedIn) publicly. Do not delay.
Jun 28 - Jul 5	Deliver V6 (“why hire me”) in at least one mock study. Get feedback. Do not tweak more than once.

After July 5, 2027 — the pitch is live. You are not rehearsing. You are shipping.

The Meta-Pitch¶

Underneath all six versions is one meta-claim: you are the engineer who noticed the atrophy, made a plan, and executed it in public on a calendar. This is rarer than any technical skill. Technical skill is legible on a resume. Self-awareness plus discipline is legible only in a story with dates.

The six versions above tell that story at six different lengths. Every version has: the atrophy noticed, the plan made, the evidence shipped, the direction stated. Keep those four beats and you can improvise the rest.

Return to README.md · Previous: 06_background_alignment.md · The fortress ends here.