Résumé guide¶
How to make the résumé survive a technical deep dive: fix the open issues first, rewrite bullets with a formula that forces evidence, and check the result against a list built for core-AI roles. Source: resume.tex (Jake Gutierrez / sb2nov template, compile on Overleaf) · current PDF: LourduRaju_Resume.pdf.
Contents: Open issues · Bullet formula · Before and after · Checklist for core-AI roles · Section order · Rules and versioning
Open issues: verify before sending anything¶
Warning
Several numbers below appeared in earlier drafts (an old story bank and feedback notes) that were written from memory, not from your records. Until you have checked each against a source you could show an interviewer (an experiment log, a report, a dashboard, an email), treat it as unverified and leave it off.
| # | claim (where it appears) | the problem | what you need before keeping it |
|---|---|---|---|
| 1 | OCR exact-match 82.74% → 97.04% (Sujanix bullet) vs 89% → 97%, a "1.2M-image" set and "~200 req/s" (old story bank, feedback) | the numbers disagree with each other | one consistent pair of numbers; the test-set size, how it was split from training data, what "exact match" means (whole string? per digit?), and who measured it |
| 2 | SVRT ("a Swin-V2-based regression transformer", Sujanix bullet) vs SVTR backbone (Transformers-OCR project) | SVTR is a published scene-text architecture (Du et al., IJCAI 2022); "SVRT" reads like a typo or a different model | intentional, consistent naming, and a 2-minute explanation of the architecture and why you chose it |
| 3 | inference cost −30% at 99.9% uptime (Sujanix bullet) | a percentage needs a baseline; uptime needs a measurement window | the before/after cost per request or per month, and where the uptime figure came from (monitoring tool, period) |
| 4 | assessment time −70% (30 min → 9 min) (ECHOME) | a time saving needs users and a method | how many sessions or users, measured how, compared with what |
| 5 | forecasting accuracy +15% (internship bullet) | "accuracy" for forecasting is ambiguous; relative or absolute? | the metric (MAPE, RMSE, …), the baseline model, the evaluation period |
| 6 | skills line: "strong CS foundations: DSA, ML system design, distributed systems", Kubernetes, SageMaker, Vertex AI, vLLM, C++, gRPC | every keyword is an interview question | keep only what you can go three "why"s deep on; move the rest out or into "exposure to" |
| 7 | experience length | "2+" overstates it | say "about 2 years" (internship, founder and Sujanix together) |
| 8 | no from-scratch or research work listed | core-AI screeners look for it first | add a Selected technical work section as labs and experiments land (examples below) |
Bullet formula¶
Verb + what you built + how (the technical decision) + measured result with context (baseline, dataset, error bar where it applies).
A number without its denominator invites the question "out of what?". Give the test-set size, the baseline and, for model quality, an interval or at least the variance across seeds.
Before and after rewrites¶
Generic examples to show the pattern. All numbers in the "after" column are illustrative; use only your own measured numbers.
| before | after (illustrative numbers) | what changed |
|---|---|---|
| Worked on LLM fine-tuning for a chatbot. | Fine-tuned a 1B-parameter model with LoRA (rank 16) on 12k support transcripts; task success on a 400-conversation held-out set rose from 61% to 74% (95% CI ±4 pts), scored by a rubric grader checked against 200 human labels. | method, data size, baseline, held-out result with an interval, how it was graded |
| Responsible for the RAG pipeline. | Rebuilt retrieval for an internal-docs assistant as BM25 + dense hybrid with reciprocal-rank fusion; recall@10 on 300 labeled queries rose from 0.62 to 0.81 with p95 latency under 800 ms. | ownership shown by a decision, a retrieval metric, a latency constraint |
| Optimized model inference. | Cut GPU cost per request 38% by moving a 7B model to INT4 weight-only quantization and continuous batching; product-eval accuracy changed −0.4 pts, within its ±0.6 CI. | the levers, the cost metric, and the quality check that makes the saving credible |
| Implemented a transformer from scratch. | Implemented a Llama-style decoder (RoPE, GQA, SwiGLU) and a Triton FlashAttention forward kernel; matches PyTorch SDPA to 1e-3 and reaches N% of its throughput at 2k context on an 8 GB laptop GPU. Code and benchmarks: link. | specific components, a correctness check, a measured comparison, a link |
| Achieved 99% accuracy on OCR. | Raised exact-match accuracy on a frozen 5,000-image held-out set of meter photos from X% to Y% by replacing a CRNN with a transformer recognizer and a class-balanced CTC loss. | the test set is named and frozen; the change that caused the gain is stated |
| Managed a team to deliver client projects. | Scoped, priced and delivered N fixed-price ML projects; brought in 3 contractors under written scopes when volume exceeded one person; |
the decisions you owned, a count, a sourced outcome |
| Proficient in PyTorch, TensorFlow, JAX, Kubernetes, Spark, CUDA, … | PyTorch (daily; wrote custom autograd functions and Triton kernels), Hugging Face, ONNX Runtime; working knowledge of Docker and AWS Lambda. | fewer tools, each with depth you can defend |
Tip
After writing each bullet, ask the three questions an interviewer will ask: How did you measure that? What was the baseline? What would you do differently? If you can't answer all three in under a minute each, rewrite or cut the bullet.
Checklist for core-AI roles¶
Run it before every version you send.
Content - [ ] The top third of the page shows the strongest core-AI evidence (from-scratch work, experiments, research-style write-ups), not the skills list. - [ ] A Selected technical work section with 3–5 items, each linking to code or a write-up. Examples as labs land: "Implemented GRPO from scratch; on a 0.5B model improved accuracy on a verifiable task from X to Y (95% CI ±Z)"; "Triton FlashAttention kernel at N% of PyTorch SDPA throughput on an RTX 5060"; merged PRs with links. - [ ] Every quality number has its dataset, its baseline and, where it applies, an interval or seed variance. - [ ] Every number is in the verified column of the open-issues table. - [ ] Each keyword survives three "why"s. - [ ] Compute is stated where it matters ("on one 8 GB GPU"): it makes results credible and shows resourcefulness. - [ ] Production experience is framed in terms core-AI teams care about: evaluation, data quality, latency and cost trade-offs, quantization, failure analysis.
Form
- [ ] One page, single column, standard section headings, text selectable in the PDF (ATS-safe; no text in images or tables).
- [ ] Links work (GitHub, write-ups, Hugging Face) and the linked repos have READMEs with results and one-command reproduction.
- [ ] Consistent tense, date format and name spelling (for example, the model name in issue 2).
- [ ] File named Firstname_Lastname_Resume.pdf; no photo, age or marital status.
Tailoring - [ ] Ordered for the target team using its company guide: kernels and inference first for an inference team, post-training and evals first for a post-training team. - [ ] The summary line matches the team, in one sentence, and says nothing you cannot link.
Section order by target¶
| target | order |
|---|---|
| frontier-lab research engineer | Selected technical work → Experience → Projects → Skills → Education |
| applied ML / LLM product roles | Experience → Selected technical work → Projects → Skills → Education |
| founder-track credibility (investors, customers) | not a résumé: a one-paragraph bio plus links to shipped systems and write-ups |
Rules and versioning¶
One page; PDF; single column. Version every change and log which version went where:
version · date · change summary · sent to (company, role, date) · outcome
v2.1 · 2027-05-04 · added GRPO result, cut skills line · <company>, RE post-training, 2027-05-06 · screen
Re-read the log before each interview so you know which claims that interviewer has seen.