sunuit.dev
8 min read
CS Internship & New-Grad Cheatsheet

Get past the bots. Land the OA.

Most resumes die in two places before a human ever reads them: the ATS that parses your file, and the recruiter who skims it for thirty seconds. This is the field guide to surviving both, and turning applications into online assessments.

Sunny Park
SWE + ML intern, ex-DoorDash · Connect on LinkedIn
01  Resume file submitted 02  ATS parses your file 03  Recruiter 30-second skim 04  Keyword match 05  Referral boost OA / SCREEN
The application funnel. Every gate removes people, mostly before anyone judges your engineering.
30s
A recruiter's first-pass read
1pg
The rule for students
5–10
Keywords searched at once
#1
Referral: the surest way in
Stage 01 · The machine reads first

Make it parse

Before a person sees anything, the ATS flattens your PDF to plain text. Fancy layouts shatter: columns interleave, text boxes vanish, tables scramble. A resume the software cannot read scores zero.

Do this0 / 6
And never
  • Multi-column templates, sidebars, text boxes. They interleave into gibberish.
  • Tables, graphics, skill bars, headshots. Read out of order, or invisible.
  • Contact info in the header or footer. Parsers skip those regions.
  • Canva or Figma "designer" resumes. Beautiful, and unparseable.
The five-second test Select all the text in your PDF, copy, paste into a blank note. If it comes out scrambled, that is what the ATS sees. Fix the layout until the paste reads top to bottom.
Stage 02 · Impact, not activity

Write bullets that convert

A human skims for evidence you did real, measurable work. Every bullet answers what you did, how (which tools), and what changed. The template underneath a good one is always the same.

Accomplished X, as measured by Y, by doing Z.

Spotlight the result
Before
Worked on a machine learning model to improve assignment decisions using company data.
After
Shipped a 2-stage ML pipeline into the core assignment system, lifting acceptance from N% to M% (Snowflake, PySpark).
Spotlight the stack
Before
Built a dine-in ordering app for restaurants using several technologies.
After
Led full-stack dev of real-time dine-in systems across iOS (Swift), Android (Kotlin), Web (Next.js), and FastAPI/PostgreSQL.
Spotlight the scale
Before
Built an NLP pipeline to find failure symptoms in service logs.
After
Engineered an NLP pipeline (KeyBERT, BERTopic) to extract failure symptoms from NM+ service logs.
Run this pass on your bullets0 / 4
No real numbers yet? One is always hiding: users served, requests per day, latency, percent faster, hours saved, records processed, coverage, team size. Estimate honestly and be ready to defend it, since every number is a fair interview question.
Stage 03 · Match the search

Get the keywords right

Recruiters search five to ten terms at a time, and ATS ranking rewards overlap with the posting. You are not gaming a robot; you are making the words you earned match the words they search.

Keyword discipline0 / 5
The highest-ROI habit Check every resume against the posting. Run it through JobRight (or Jobscan) to score your match and surface missing keywords, then submit a customized resume for each job. Tailoring the top third and keywords per posting beats everything else. Do not fire the same PDF at two hundred roles.
The real edges · where most people plateau

The quieter craft

The same experience can read as SWE or MLE, but a resume that tries to be both reads as neither.

Pick one lens and make the whole page reinforce it: which role leads, which bullets rise, which stack you emphasize, even which projects you keep. Applying to two kinds of role? Keep two versions.

My own work, each project read two ways

An ML system serving live assignment decisions
Read as SWEShipped decline-reason inference into a core assignment system with model monitoring and dashboards, serving decisions in production.
Read as MLETrained a dual-objective LightGBM (N AUC-PR) on NM+ offers, soft-labeled with an LLM and distilled into the served model.
A real-time platform I co-founded
Read as SWEArchitected a multi-tenant real-time platform (Next.js, FastAPI, Postgres, WebSockets) on AWS, syncing state across iOS, Android, and web.
Read as MLEBuilt a demand-pacing model (LightGBM) on historical usage data, cutting wait times and improving throughput.

Same underlying work, framed for the role. The SWE reads foreground systems, serving, and infra; the MLE reads foreground data, models, and metrics. A resume commits to one of these; it does not try to be both.

Order tells a story: my DoorDash role, as it runs

  1. Developed 2-stage ML pipeline to infer Dasher decline reasons from NM+ delivery offers (Snowflake, PySpark).
  2. Benchmarked contrastive embeddings (HDBSCAN/k-means) against archetypal analysis for decline reason discovery.
  3. Modeled declines as blends of N continuous reasons; soft-labeled with Sonnet 4.5 for distillation into the served model.
  4. Trained dual-objective LightGBM (N AUC-PR decline gate, M top-1 reason head) with SHAP interpretability.
  5. Deployed into ____, DoorDash's core assignment system, boosting Dasher acceptance rate from N% to M%.

The highlighted N and M stand in for my real figures, and the internal system name is blanked, so the specifics stay mine.

This order builds an arc: the problem and data, the method, the model and its metrics, then the deployed business result as the payoff. The one rule, whichever way you go, is to never bury your strongest result in the middle. Leading with it suits skim-heavy recruiters; ending on it, as here, rewards anyone who reads the whole role.

Then order by strength0 / 5
The move is the same in both Lead with the metric, name the real system, weave the stack into the sentence, vary your verbs, and list only skills you can defend. What changes per lens is which of those you put first, and which numbers you reach for.
Stage 04 · Get found, not just seen

Treat LinkedIn as SEO

Recruiters search LinkedIn the way they search the ATS. Set it up once, correctly, and you start getting inbound: the rare messages that skip the queue.

Profile setup0 / 5
Why it works A recruiter filtering candidates sees the same keyword-and-skills surface your ATS resume shows. A high match rate on a public profile is what turns a cold search into an inbound message, months before you would have found the role yourself.
Stage 05 · The highest-leverage move

Referrals are the cheat code

A referral is the single biggest jump in your odds of a screen. It routes you around the cold pile and puts a name behind your file. Not all are equal: the closer the referrer to the decision, the more their word is worth.

1
Executive or senior IC you worked with
Speaks first-hand to your work. Most weight.
Strongest
2
Management, directors
Their word reaches hiring managers.
Strong
3
Functional manager on the team
Runs the team you would join.
Good
4
Teammates and peers
The people you'd work with daily.
Solid
5
An adjacent colleague
Near the team, not on it. Still beats none.
Weakest
Where to find one0 / 5
Hi [Name], I'm a CS student at [School] and admire the [team/product] work at [Company]. I'm applying for the [role] and would love a quick read on whether I'm a fit. Open to a referral? Resume attached. Thank you!
Stage 06 · Timing and volume

Apply like a system

Recruiting is a race, not a deadline. Roles open early and close when the pipeline fills.

  • Apply early and fast. Postings open Aug to Oct, fill rolling
  • Customize per posting, batch the rest. JobRight tailors the top third
  • Target 20 to 40 companies with alerts; track everything
  • Volume with a referral beats volume alone
Stage 07 · Where this ends

The OA gate

Do the above and OAs hit your inbox: timed tests on HackerRank, CodeSignal, Codility. The last gate before a human, and pure prep.

  • NeetCode 150 / Blind 75 by pattern, not grinding
  • Standard library cold: heaps, maps, sorting, strings
  • Practice timed. The clock is what fails people
In the test Read constraints first (n hints the complexity). Get brute force passing, then optimize. Test edges. Never leave it blank; partial credit is per test case.
Pre-flight · before every submit

The ATS-ready checklist

One last gate. Run this on every resume before it goes out. Your ticks save in this browser.

Pre-flight0 / 16
The toolkit

What I actually use

The tools that run the funnel. Most have a free tier. They speed up the grind; they do not replace a strong resume or a real referral.

JobRightCheck · Customize · Refer
Scores your resume against each posting, generates a customized version per job, matches roles, and surfaces referrals. Check against each posting; submit a tailored resume every time.
Google DocsResume · Draft
Where I draft and iterate on the resume, easy to share and collect comments. Overleaf (LaTeX) is my preferred option when I want a pixel-perfect, ATS-safe layout.
NeetCode / LeetCodeOA · Interview Prep
Learn the patterns on NeetCode (150, Blind 75), then grind and time yourself on LeetCode. OA prep that builds recognition, not memorization.
HandshakeCampus Recruiting
The university portal for intern and new-grad roles and events. Early roles and recruiter events aimed at students.