06 — The M13 Pitch¶

By M13 you have two capstones live, a blog, an overhauled LinkedIn, and a resume. What you still need is the sentence — the one you can say without notes, at any moment, that makes the person on the other end lean in. This file is that sentence, in six shapes for six audiences.

Write your six versions in W53, tape them to the wall, and rehearse them for two weeks before the first recruiter call in W55. The point is not memorization — it is fluency. You should be able to open with any of these mid-Zoom without your throat closing.

The building blocks¶

Every version below is assembled from the same five artifacts. Reference them by their short handles:

P5.3 — offline ONNX inference CLI with benchmarks (Phase 5 capstone)
P6.1 — MiniServe precursor: gRPC ML inference service with observability
Capstone Alpha (MiniServe) — production-shaped C++ inference server, gRPC + REST + dynamic batching + OTel, benchmarked vs FastAPI+PyTorch
Capstone Beta (PyFast) — native Python extension on PyPI (recommended: pyfast_ann, HNSW-based ANN)
The profiling case study blog post — real hotspot, real >2× speedup, real flamegraphs

Every pitch names at least one of these with a number attached. No number = no pitch.

Version 1 — The 30-second elevator¶

Use when: someone at a meetup, a Zoho hallway, an airport lounge asks “so what do you do?” This is the tightest version. Practice until it comes out in one breath.

I’m an applied ML engineer at Zoho — the last year I’ve been going deep on C++ for ML infrastructure. I shipped a gRPC-based ML inference server in C++ that runs {X}× faster than the FastAPI+PyTorch baseline on the same hardware, and I open-sourced a Python-native ANN library on PyPI that binds to a C++ HNSW core. Right now I’m looking for teams building ML platform or inference infrastructure where the C++ layer actually matters.

Sixty to seventy words. Ends with the ask. Do not extend it unless they ask a follow-up — and they will.

Version 2 — The 2-minute recruiter call intro¶

Use when: a recruiter opens with “tell me about yourself” on the discovery call. This is the version they will paraphrase into the hiring manager’s inbox — so make it easy to paraphrase.

Sure. I’ve been an applied ML engineer at Zoho for the last {N} years, working on {one-line Zoho ML work}. Over the last year I ran a personal deep-dive into C++ for ML systems — not casually, but as a full curriculum: memory management, concurrency, network programming, and production observability.

The output of that work is two open-source artifacts I ship as my portfolio. The first is MiniServe — a C++ ML inference server exposing gRPC and REST, with dynamic batching, Prometheus metrics, and OpenTelemetry tracing. I benchmarked it against a FastAPI+PyTorch equivalent and it holds {X}× throughput and {Y}ms tighter p99 latency on the same hardware. The second is a Python-native ANN library on PyPI — nanobind bindings around a C++ HNSW core — that gives numpy-native ergonomics with C++ speed.

On top of that I write a blog on the same domain — the most-read post is a profiling case study where I took a slow function and got a {Z}× speedup, with flamegraphs before and after.

What I’m looking for next is a team where the C++/ML boundary is where the interesting problems live — ML platform, inference infra, or a data-intensive product with real perf constraints. I’m open to India-based, India-hybrid, or India-remote for global teams.

Two minutes exactly at natural pace. Ends with what you want next. The recruiter now has three sound bites to pitch: MiniServe, PyFast, blog.

Version 3 — The 5-minute technical intro (start of a technical screen)¶

Use when: the study partner opens with “walk me through your background before we dive in”. Do not read a resume — narrate a career, then land on the artifact you most want to discuss.

I trained as {your training}, joined Zoho in {year} on {team}, and shifted into applied ML around {year} — mostly {tasks: model training, feature engineering, deployment on Catalyst}.

The gap I kept seeing on my team was that our ML services were all Python — which is fine until you need to serve them at low latency, or you need to embed a model inside another service, or you need to profile something that isn’t just “the GIL”. So starting in {2026-07}, I ran a personal 13-month program to build applied C++ depth for ML systems. It covered the language itself through C++20, then memory and concurrency, then network programming, then production systems — gRPC, observability, profiling, allocators, compiler optimization.

The two things I’d point you to are MiniServe and PyFast — I can go deep on either. MiniServe is the production side: C++ inference server with dynamic batching, gRPC + REST, OTel tracing, Docker-deployed, benchmarked. PyFast is the library side: nanobind-based HNSW ANN index on PyPI, numpy-native API, {X}× faster than {baseline} on the {benchmark} dataset.

The tradeoff I made repeatedly through the program is: fewer artifacts, deeper artifacts. Two things shipped and benchmarked, not ten things half-done. Happy to start wherever you want — the code and benchmarks are open.

Five minutes at natural pace. Ends with a hand-off. The study partner now picks the artifact, and you are on your home turf.

Version 4 — Portfolio landing page hero (written, not spoken)¶

Use when: your personal domain’s root page. This is the version a recruiter or hiring manager reads before deciding whether to click through.

Raghul {surname} — applied ML engineer working at the C++/ML boundary.

I build the infrastructure that ships ML models to production. Two artifacts I’d point you to:

MiniServe — a C++ ML inference server. gRPC + REST, dynamic batching, Prometheus + OpenTelemetry, Docker-deployed. Benchmarked at {X} req/s p99 {Y}ms — {Z}× the throughput of an equivalent FastAPI+PyTorch server on the same hardware. Repo · Blog post · Live demo

PyFast — a Python-native ANN library on PyPI. C++ HNSW core with nanobind bindings; numpy-native API. {A}× faster than {baseline} on {benchmark}. PyPI · Repo · Notebook

I’m based in {city}, India. Currently at Zoho. Open to conversations about ML platform, inference infrastructure, and roles where C++ is the wedge.

Resume PDF · LinkedIn · Email

Above the fold. No animations, no scroll-jacking, no “hero video”. A recruiter should be able to answer “should I forward this to the hiring manager?” in 15 seconds. Give them the numbers they need to say yes.

Version 5 — LinkedIn headline + short bio¶

Use when: everyone who sees your profile. This is the version that runs 24/7 without you present.

Headline (220 char limit; keep under 180 for mobile):

Applied ML Engineer · C++ for ML infrastructure · built MiniServe (C++ inference server) + PyFast (native ANN on PyPI)

About — three paragraphs, ~180 words total:

I’m an applied ML engineer with a growing specialty in the C++/ML boundary — the layer where Python model code meets production latency, throughput, and observability constraints. Based in {city}, currently at Zoho.

Over the last year I shipped two artifacts that anchor my portfolio. MiniServe is a C++ ML inference server (gRPC + REST, dynamic batching, Prometheus + OpenTelemetry) benchmarked at {X}× the throughput of a FastAPI+PyTorch equivalent on the same hardware. PyFast is a nanobind-based Python extension on PyPI that gives numpy-native ergonomics to a C++ HNSW ANN core. Both open-source, both documented, both with reproducible benchmarks.

Next: ML platform, inference infrastructure, or data-intensive product teams where the C++ layer is the wedge. Open to India-based, India-hybrid, or India-remote for global teams. The fastest way to see how I work is the profiling case study on my blog — real hotspot, {Z}× speedup, before/after flamegraphs.

Version 6 — “Why should I hire you?” (hiring manager close)¶

Use when: end of an onsite loop, hiring manager 1:1, or the “any final questions?” flip. Do not answer this generically. Answer it with the specific role in mind — but the skeleton is the same.

Three reasons that I think are hard to fake.

First — I already ship the thing you are hiring for. MiniServe is not a toy; it has the same shape as {their inference service / their platform / their tool}, at smaller scale. The learning curve on your codebase will be shorter than for a candidate who has to acquire the shape from scratch.

Second — I measure. Every artifact in my portfolio has a benchmark and a methodology paragraph. The profiling case study on my blog is the clearest signal: I found a real hotspot, applied a specific fix, and re-measured with numbers. On your team I will do the same before touching latency-critical code — and I will disagree with someone who ships a “faster” change with no measurement.

Third — I taught myself the C++ layer while working full-time in ML. That did not come from a course; it came from thirteen months of shipping small things. Whatever the next thing your team needs, I will learn it the same way.

What would make me a wrong hire — happy to talk about that too if it helps.

Ends by inviting disqualification. study partners remember candidates who ask “why wouldn’t you hire me?” — because most don’t.

Rehearsal schedule¶

Fluency is a physical skill, not an intellectual one. Rehearse out loud, not in your head.

Week	What to rehearse	How
W53	V1 (30-sec)	20 reps into your phone recorder over 5 days
W53	V4 + V5 (written)	Ship them, then read them aloud once daily
W54	V2 (2-min recruiter)	10 reps, timed, into recorder. Play back once. Do not cringe — take notes.
W54	V3 (5-min technical)	5 reps with a peer or in front of a mirror. Peer beats mirror.
W55	V6 (why-hire-you)	5 reps, but customize for 3 of your Tier A targets specifically
W55-56	All six	One pass through all six every Monday morning as warm-up

Do not rehearse the day of a real call. That is the day for tea, deep breaths, and re-reading the JD.

What NOT to claim¶

The pitch works only if every claim in it is defensible. One inflated line and the study partner will pull the thread, and the whole thing unravels. Bright red lines you will not cross:

Do not claim MiniServe production traffic you don’t have. “Ran a load test at {X}” is honest. “Serves {X} req/s in production” is not, unless it does.
Do not describe the 13-month program as “systems engineering experience” as if it were a job. It is a portfolio. Call it what it is.
Do not compare yourself to Triton, Ray Serve, or vLLM without qualifiers. “MiniServe is a smaller-scale, single-node version of the pattern Triton uses for dynamic batching” is fine. “MiniServe is a Triton alternative” is not.
Do not benchmark against straw men. If your FastAPI+PyTorch baseline is unoptimized, say so and give the study partner the version number and config. A hiring manager who spots an unfair benchmark once will discount everything you say after.
Do not claim PyPI download counts as a proxy for adoption in your first two months. Downloads on a new package are mostly you and mirror bots. Mention adoption only when you have real external stars and issues.

The pitch is strongest when every sentence in it is a sentence you would defend in a deposition. Build it that way.

The one-sentence version¶

For the days you have no time to think — memorize this as your default:

I ship applied C++ for ML systems — I built a production-shape inference server and a native Python ANN library, both open-source and benchmarked, and I’m looking for teams where that combination is the wedge.

Thirty-eight words. Say it once a day for two weeks. It will come out of your mouth without effort by W55.

Nav: ← 05 study Conversion · → projects.md · Phase 7 README · Roadmap root