# Blog — Tanishq Bhosale

> Engineering write-ups on building production LLM systems: grounding and
> anti-fabrication, verify-and-repair loops, LLM evaluation, and self-hosted
> AI infrastructure. Written by Tanishq Bhosale, Full Stack & AI Engineer,
> Bangalore, India.

## How to read this site

Every post exists in two formats with identical content:

- **HTML** — `https://thetan.in/blog/<slug>` — the canonical URL. Cite this one.
- **Markdown** — `https://thetan.in/blog/<slug>.md` — same text, no markup or
  interactive widgets. Cheaper to parse and to tokenise. It is marked
  `noindex` because it duplicates the HTML page; that is a search-listing
  directive only, and it is free to fetch, read and quote.

The canonical HTML URL also answers with Markdown if you send
`Accept: text/markdown`.

Related: https://thetan.in/llms.txt (author, skills, experience, contact).

## Posts

### Your AI Resume Builder Is Lying About You

I built ResumeOS to tailor a resume to any job description without inventing a single fact about you. This is the grounding architecture that makes that a guarantee instead of a hope — a judge model handed the whole profile as ground truth, a bounded verify-and-repair loop, and one rule: fail closed on correctness, fail open on cosmetics.

- Canonical page (HTML): https://thetan.in/blog/resume-builder
- Full text (Markdown): https://thetan.in/blog/resume-builder.md
- Published: 2026-07-27 · Updated: 2026-07-27
- Reading time: ~14 min
- Topics: LLM Engineering, Guardrails, FastAPI, LaTeX, Prompt Injection, Self-hosting
- About the project: ResumeOS — https://resume.thetan.in

Key takeaways:
- AI resume builders fabricate skills because nothing in 'close the gap between this resume and this job' distinguishes repositioning real experience from inventing new experience.
- ResumeOS grounds every generated resume against a judge model that receives the entire profile JSON as the allowed-fact set and must quote any offending claim verbatim.
- Positioning is free, facts are not: paraphrase, reordering and repositioning all pass; a skill, employer, metric or date that appears nowhere in the profile is a violation.
- A grounding check is only as good as the ground truth you hand it — passing a lossy profile extract made the judge flag 16 of 16 of a user's real facts as fabrications.
- Fail closed on correctness, fail open on cosmetics: refuse to ship a resume that will not compile or still contains a fabricated fact; ship a slightly-over-one-page resume with a warning.
- Cosine similarity is the wrong tool for a job-fit score — it measures vocabulary overlap rather than capability, and is least reliable in the middle of the range where the pass/fail threshold sits.
- Model-produced edits are applied as {find, replace} pairs only when find occurs exactly once, so an edit can never silently corrupt the document.

Questions this post answers:
- What is ResumeOS?
- How does ResumeOS stop AI from fabricating skills on a resume?
- Why do AI resume builders add skills you don't have?
- What is a verify and repair loop in an LLM pipeline?
- Should you use embeddings for job description matching?
- What makes a resume ATS-friendly?
- How do you protect an LLM app from prompt injection in user input?
- What stack does ResumeOS run on?

## Author

Tanishq Bhosale — Full Stack & AI Engineer in Bangalore, India. Builds
production LLM and agentic AI systems with LangGraph, FastAPI, React and
React Native. Contact: bhosaletanishq4@gmail.com · https://thetan.in